The Analysis of Firms and Employees: Quantitative and Qualitative Approaches 9780226042893

Citation preview

The Analysis of Firms and Employees

A National Bureau of Economic Research Conference Report

The Analysis of Firms and Employees Quantitative and Qualitative Approaches

Edited by

Stefan Bender, Julia Lane, Kathryn Shaw, Fredrik Andersson, and Till von Wachter

The University of Chicago Press Chicago and London

STEFAN BENDER is a senior researcher at the Institute for Employment Research. JULIA LANE is a senior vice president of Economics, Labor, and Population Studies at the National Opinion Research Center at the University of Chicago and a senior research fellow at the U.S. Bureau of the Census. KATHRYN SHAW is the Ernest C. Arbuckle Professor of Economics in the Graduate School of Business, Stanford University, and a research associate of the National Bureau of Economic Research. FREDRIK ANDERSSON is a senior research associate of the Cornell Institute for Social and Economic Research and a research fellow with the Longitudinal Employer-Household Dynamics Program (LEHD), U.S. Bureau of the Census. TILL VON WACHTER is assistant professor of economics at Columbia University and a faculty research fellow of the National Bureau of Economic Research.

The University of Chicago Press, Chicago 60637 The University of Chicago Press, Ltd., London © 2008 by the National Bureau of Economic Research All rights reserved. Published 2008 Printed in the United States of America 17 16 15 14 13 12 11 10 09 08 1 2 3 4 5 ISBN-13: 978-0-226-04287-9 (cloth) ISBN-10: 0-226-04287-1 (cloth)

Library of Congress Cataloging-in-Publication Data The analysis of firms and employees : quantitative and qualitative approaches / edited by Stefan Bender . . . [et al.]. p. cm. — (A National Bureau of Economic Research conference report) Includes bibliograpical references and index. ISBN-13: 978-0-226-04287-9 (cloth : alk. paper) ISBN-10: 0-226-04287-1 (cloth : alk. paper) 1. Employees. 2. Industrial organization. 3. Industrial sociology. I. Bender, Stefan. HD4901.A652 2008 331.1—dc22 2008008259

o The paper used in this publication meets the minimum requirements of the American National Standard for Information Sciences— Permanence of Paper for Printed Library Materials, ANSI Z39.48-1992.

National Bureau of Economic Research Officers Elizabeth E. Bailey, chairman John S. Clarkeson, vice-chairman Martin Feldstein, president and chief executive officer Susan Colligan, vice president for administration and budget and corporate secretary

Robert Mednick, treasurer Kelly Horak, controller and assistant corporate secretary Gerardine Johnson, assistant corporate secretary

Directors at Large Peter C. Aldrich Elizabeth E. Bailey John H. Biggs John S. Clarkeson Don R. Conlan Kathleen B. Cooper Charles H. Dallara George C. Eads

Jessica P. Einhorn Martin Feldstein Roger W. Ferguson, Jr. Jacob A. Frenkel Judith M. Gueron Robert S. Hamada Karen N. Horn John Lipsky

Laurence H. Meyer Michael H. Moskow Alicia H. Munnell Rudolph A. Oswald Robert T. Parry Marina v. N. Whitman Martin B. Zimmerman

Directors by University Appointment George Akerlof, California, Berkeley Jagdish Bhagwati, Columbia Ray C. Fair, Yale Michael J. Brennan, California, Los Angeles Glen G. Cain, Wisconsin Franklin Fisher, Massachusetts Institute of Technology Saul H. Hymans, Michigan Marjorie B. McElroy, Duke

Joel Mokyr, Northwestern Andrew Postlewaite, Pennsylvania Uwe E. Reinhardt, Princeton Nathan Rosenberg, Stanford Craig Swan, Minnesota David B. Yoffie, Harvard Arnold Zellner (Director Emeritus), Chicago

Directors by Appointment of Other Organizations Richard B. Berner, National Association for Business Economics Gail D. Fosler, The Conference Board Martin Gruber, American Finance Association Arthur B. Kennickell, American Statistical Association Thea Lee, American Federation of Labor and Congress of Industrial Organizations William W. Lewis, Committee for Economic Development

Robert Mednick, American Institute of Certified Public Accountants Angelo Melino, Canadian Economics Association Jeffrey M. Perloff, American Agricultural Economics Association John J. Siegfried, American Economic Association Gavin Wright, Economic History Association

Directors Emeriti Andrew Brimmer Carl F. Christ George Hatsopoulos Lawrence R. Klein

Franklin A. Lindsay Paul W. McCracken Peter G. Peterson Richard N. Rosett

Eli Shapiro Arnold Zellner

Relation of the Directors to the Work and Publications of the National Bureau of Economic Research 1. The object of the NBER is to ascertain and present to the economics profession, and to the public more generally, important economic facts and their interpretation in a scientific manner without policy recommendations. The Board of Directors is charged with the responsibility of ensuring that the work of the NBER is carried on in strict conformity with this object. 2. The President shall establish an internal review process to ensure that book manuscripts proposed for publication DO NOT contain policy recommendations. This shall apply both to the proceedings of conferences and to manuscripts by a single author or by one or more coauthors but shall not apply to authors of comments at NBER conferences who are not NBER affiliates. 3. No book manuscript reporting research shall be published by the NBER until the President has sent to each member of the Board a notice that a manuscript is recommended for publication and that in the President’s opinion it is suitable for publication in accordance with the above principles of the NBER. Such notification will include a table of contents and an abstract or summary of the manuscript’s content, a list of contributors if applicable, and a response form for use by Directors who desire a copy of the manuscript for review. Each manuscript shall contain a summary drawing attention to the nature and treatment of the problem studied and the main conclusions reached. 4. No volume shall be published until forty-five days have elapsed from the above notification of intention to publish it. During this period a copy shall be sent to any Director requesting it, and if any Director objects to publication on the grounds that the manuscript contains policy recommendations, the objection will be presented to the author(s) or editor(s). In case of dispute, all members of the Board shall be notified, and the President shall appoint an ad hoc committee of the Board to decide the matter; thirty days additional shall be granted for this purpose. 5. The President shall present annually to the Board a report describing the internal manuscript review process, any objections made by Directors before publication or by anyone after publication, any disputes about such matters, and how they were handled. 6. Publications of the NBER issued for informational purposes concerning the work of the Bureau, or issued to inform the public of the activities at the Bureau, including but not limited to the NBER Digest and Reporter, shall be consistent with the object stated in paragraph 1. They shall contain a specific disclaimer noting that they have not passed through the review procedures required in this resolution. The Executive Committee of the Board is charged with the review of all such publications from time to time. 7. NBER working papers and manuscripts distributed on the Bureau’s web site are not deemed to be publications for the purpose of this resolution, but they shall be consistent with the object stated in paragraph 1. Working papers shall contain a specific disclaimer noting that they have not passed through the review procedures required in this resolution. The NBER’s web site shall contain a similar disclaimer. The President shall establish an internal review process to ensure that the working papers and the web site do not contain policy recommendations, and shall report annually to the Board on this process and any concerns raised in connection with it. 8. Unless otherwise determined by the Board or exempted by the terms of paragraphs 6 and 7, a copy of this resolution shall be printed in each NBER publication as described in paragraph 2 above.

Contents

Acknowledgments

ix

Introduction Stefan Bender, Julia Lane, Kathryn Shaw, Fredrik Andersson, and Till von Wachter

1

I. HUMAN RESOURCE PRACTICES AND FIRM PRODUCTIVITY 1. The Effect of HRM Practices and R&D Investment on Worker Productivity Fredrik Andersson, Clair Brown, Benjamin Campbell, Hyowook Chiang, and Yooki Park 2. Using Behavioral Economic Field Experiments at a Firm: The Context and Design of the Truckers and Turnover Project Stephen V. Burks, Jeffrey Carpenter, Lorenz Götte, Kristen Monaco, Kay Porter, and Aldo Rustichini 3. Subjective Evaluation of Performance and Evaluation Interview: Empirical Evidence from France Marc-Arthur Diaye, Nathalie Greenan, and Michal W. Urdanivia

19

45

107

vii

viii

Contents

II. FIRM DIFFERENCES IN HUMAN RESOURCES PRACTICES 4. Do Initial Conditions Persist between Firms? An Analysis of Firm-Entry Cohort Effects and Job Losers Using Matched Employer-Employee Data Till von Wachter and Stefan Bender

135

5. Changes in Workplace Segregation in the United States between 1990 and 2000: Evidence from Matched Employer-Employee Data Judith Hellerstein, David Neumark, and Melissa McInerney

163

6. The Effect of Tuition Reimbursement on Turnover: A Case Study Analysis Colleen Flaherty Manchester

197

III. EFFECTS OF OWNERSHIP CHANGES ON THE ORGANIZATION OF PRODUCTION 7. Ownership and Wages: Estimating Public-Private and Foreign-Domestic Differentials with LEED from Hungary, 1986 to 2003 John S. Earle and Álmos Telegdy 8. Insider Privatization and Careers: A Study of a Russian Firm in Transition Guido Friebel and Elena Panova

229

253

IV. GLOBALIZATION, TRADE, AND LABOR MARKETS 9. Trade and Workforce Changeover in Brazil Marc-Andreas Muendler

269

10. Job Creation Abroad and Worker Retention at Home Sascha O. Becker and Marc-Andreas Muendler

309

11. Wage and Productivity Premiums in Sub-Saharan Africa Johannes Van Biesebroeck

345

Contributors Author Index Subject Index

373 377 383

Acknowledgments

We would like to acknowledge the substantial contribution of a number of people whose contribution was critical to the success of this book and of the original Conference on the Analysis of Firms and Employees (CAFE) 2006 in Nuremberg, Germany. The sponsors of the conference were critical to its success. We thank Gail Pesyna of the Sloan Foundation; the Economics Program, the Decision, Risk and Management Sciences Program, and the Sociology Program of the National Science Foundation (NSF); the Deutsche Forschungsgemeinschaft (German Research Foundation); and Bernd Fitzenberger of the Research Network “Flexibility in Heterogeneous Labour Markets” of the German Research Foundation and the Institute for Employment Research (IAB). We also thank the Research Data Center staff in Nuremberg who worked so hard to coordinate all the conference logistics, particularly Dagmar Herrlinger and Michael Stops. We also thank John Haltiwanger for his participation in the planning meetings for the conference as well as his assistance in obtaining NSF support for the conference. Helena Fitz-Patrick of the National Bureau of Economic Research (NBER) and Parker Smathers of the University of Chicago Press provided invaluable support in getting the papers finalized and in appropriate publication format. Finally, we are grateful to all the attendees, discussants, and chairs whose thoughtful participation made the entire conference an extremely rewarding endeavor.

ix

Introduction Stefan Bender, Julia Lane, Kathryn Shaw, Fredrik Andersson, and Till von Wachter

Motivation for Book The analysis of the interactions of firms and employees has followed two distinct paths. One path has focused on large-scale, often nationally representative, data sets on firms and employees, typically housed at federal statistical agencies. In some cases, this path has intensively used administrative data, alone or integrated with survey data and, in other cases, the use of surveys designed to collect information about both firms and workers. The other path has been the development of specialized surveys and gathering of personnel records of a small number of firms (or even one firm) or intensive observation (essentially collection of qualitative data) from case studies based on site visits to firms by researchers, data typically housed at universities or think-tanks. Each of these two study approaches has uncovered interesting and useful pieces of information. Researchers working with large-scale, national, firm-level matched employer-employee data sets have begun to address a variety of organizational topics, such as determinants of wage inequality, the use of alternative wage policies (such as the use of incentive pay) and their impact on worker selection, gender differences in promotion, and Stefan Bender is a senior researcher at the Institute for Employment Research. Julia Lane is senior vice president of Economics, Labor, and Population Studies at the National Opinion Research Center at the University of Chicago, and a senior research fellow at the U.S. Bureau of the Census. Kathryn Shaw is the Ernest C. Arbuckle Professor of Economics in the Graduate School of Business, Stanford University, and a research associate of the National Bureau of Economic Research. Fredrik Andersson is a senior research associate of the Cornell Institute for Social and Economic Research, and a research fellow with the Longitudinal Employer-Household Dynamics Program (LEHD), U.S. Bureau of the Census. Till von Wachter is an assistant professor of economics at Columbia University, and a faculty research fellow of the National Bureau of Economic Research.

1

2

Stefan Bender et al.

differences in alternative career paths within and across firms. They have also been able to examine the impact of job and worker reallocation on worker and firm outcomes. Still, one weakness of existing empirical analysis of firm outcomes from large, national data sets is that the underlying determinants of firm performance are generally unknown. In contrast, researchers who have been engaged in intensive study of a small number of firms through either case studies or specialized surveys have been able to describe rich contextual variation in organizational decision making. Many of these studies have been conducted within the context of projects affiliated with the Sloan Foundation’s Industry Centers and the NBER/Sloan joint projects sponsoring site visits. These studies have highlighted, among other things, the importance of human resource practices such as the use of teams for the successful adoption for the use of new technologies (e.g., information technology) and have as a common theme the link between business success how businesses organize their workers. While such work has demonstrated the tremendous value added of specialized surveys and the insights to be derived from the intensive qualitative data collection associated with site visits, questions are often raised about the ability to generalize results from small, potentially nonrepresentative samples. The combination of these two empirical approaches means that it is possible to envision the development of a new field of economics, one that is at the nexus of labor economics, industry studies, and industrial organization. The focus of this field is inherently the organizational structure of businesses with a focus on how workers are organized. Matched employeremployee data that include the information from large-scale data sets as well as from specialized surveys and site visits provide a momentous opportunity for a research agenda that is focused on the study and understanding of the interaction of firms and workers. This book results from a conference that was planned to foster just such a research agenda.1 An overarching goal of the conference was to bring together both senior and junior researchers from the two study fields—traditional labor economists and industry studies researchers, particularly those who have conducted case studies—to illustrate the different insights to be gained from the two approaches and provide the stimulus for a next 1. The Conference on Firm and Employees (CAFE) was held September 29 to 30, 2006, in Nuremberg, Germany, sponsored by the Institute for Employment Research (IAB), the Data Access Center (FDZ-BA/IAB), The Deutsche Forschungsgemeinschaft (German Research Foundation), their Research Network “Flexibility in Heterogeneous Labour Markets,” the National Bureau of Economic Research, the Alfred P. Sloan Foundation, and the National Science Foundation. Invited keynote speakers included Dan Hamermesh (University of Texas, Austin), Ed Lazear (Stanford University), Richard Freeman (Harvard University), and Mari Sako (Oxford University), who all have done path-breaking work related to the topics of the conference. Over 160 papers were submitted, of which about 40 were accepted after an extensive refereeing process. Over 100 researchers from around the world participated.

Introduction

3

generation of research. The potential for such data to answer key empirical questions in economics has been well described elsewhere.2 The list includes the effect of firm policies on the job ladders and career paths of workers; the effects of workforce composition on business growth and survival; as well as the micro-level analysis of the demand for labor, including the impact of technological and structural change. The book features eleven papers selected from that conference. They were selected by referees on the basis of their quality as well as for the new insights that they provided about the interactions between firms and their employees. The next sections of this introduction provide an overview of the major findings that have been made possible as a result of these new data sets. These span several different facets of the relationship between firms and workers, beginning with shedding more light on the relationship between human resource practices and productivity, then examining how firm differences in the organization of production are related to differences in human resource practices, how changing ownership affects the organization of production, and, finally, how the changing trade patterns, particularly globalization, affect firm competitiveness and then works through to their employees. The final section provides an overview that highlights the innovative nature of the data sets themselves. The Major Findings Human Resource Practices and Firm Productivity Personnel economics has long been concerned with the fundamental question whether human resource practices such as incentive contracts or monitoring affect workers’ productivity and worker turnover and how they relate to firms’ efforts to innovate and gain competitiveness in an evolving marketplace. A recent surge in new data sets and data collection efforts has led to an increasing amount of ambitious empirical work describing and testing some of the key relationships between firms’ personnel strategies and worker and firm outcomes. The first three chapters in this volume give an excellent introduction into three of the most prominent and promising directions in this growing research area. Each of the three chapters is based on a unique new data source and addresses a core relationship between human resource practices at a different level of aggregation. The first chapter uses a large matched administrative employer-employee data set from the United States to analyze the relationship between human resource practices, research and development (R&D), and worker productivity in a sample of firms in the electronics industry. The second chapter augments a 2. See, for example, Abowd, Haltiwanger, and Lane (2004) and Brown, Haltiwanger, and Lane (2006).

4

Stefan Bender et al.

traditional analysis of a large firm’s personnel records with innovative survey data on workers’ preferences, attitudes, and behaviors in a social context to study the determinants of turnover and worker productivity in the trucking industry. The third chapter tests predictions of a model of subjective evaluation and effort in a moral hazard setting using data from an exceptionally detailed matched worker firm survey from the French manufacturing industry. Firms constantly face the problem of adjusting their production processes and their workforce to impulses from technological progress and increasing competition. A recurring theme is the question of how firms adjust their human resource (HR) practices to cope with the evolving economic environment, and whether some HR practices are more helpful in successfully implementing or developing new technologies than others. A core difficulty in providing an empirical answer to this question is that most data sets that contain information on investment into new technologies and R&D for a sufficient number of firms typically have little information on firms’ HR structures. The first innovation of the first chapter, by Andersson, Brown, Campbell, Chiang, and Park, is to construct measures of HR practices based on longitudinal earnings and turnover information from the universe of workers in a large sample of firms in a particular industry (electronics). The chapter then analyzes the joint of occurrence of indicators such as accession and separation rates or within-job wage growth in HR “clusters” and describes the practices of firms with high and low investment in R&D. In a last step, the authors examine whether the interaction between R&D and HR practices significantly affects worker productivity. The chapter then interprets its tremendous amount of new information in the context of an economic model where firms have to decide whether to produce technology (R&D) in house or acquire it in the market and have to structure their HR practices to train, retain, or hire the appropriate workforce. The chapter’s comprehensive descriptive empirical approach based on explicit firm-level HR measures grounded in economic intuition should pave the way to further fruitful analysis of the incidence and effect of HR practices using increasingly available matched administrative employer-employee data. One of the great benefits of this approach is the potential to analyze the personnel choices and their correlation with worker and firm outcomes for a broad range of firms, workers, and phenomena of interest. The price to pay for this gain in insight is the focus on broader measures of HR practices. While ideal for describing recurring patterns and correlations, sometimes the relationships between firms’ internal institutions and workers’ incentives and productivity emphasized by the theory are more subtle. This is especially true for more recent modeling approaches emphasizing behavioral aspects of the firm-worker relationship, such as trust or loyalty. The second chapter in this volume, by Burks, Carpenter, Götte, Monaco,

Introduction

5

Porter, and Rustichini, describes an ambitious data collection effort and research agenda aimed at uncovering specific and intricate links between HR practices, worker incentives and attitudes, and worker turnover and productivity. To do so, the chapter combines longitudinal personnel records of a large trucking firm in the United States with a panel survey of a cohort of newly hired workers geared to elicit a detailed range of behavioral and preference parameters. Thereby, a key innovation is to obtain information on risk and loss aversion or cooperation through standard survey questions as well as responses to small laboratory experiments such as prisoners’ dilemma games. The second chapter describes this tremendous effort in detail and puts it into the context of a statistical analysis of how turnover and effort evolve with tenure at the trucking firm. The results show a large and increasing amount of variance in productivity of truck drivers and a differential effect of selective exit on variance as job tenure increases. While traditional analyses of firms’ personnel records typically have to stop at this point, the added survey data will enable the authors to draw more specific conclusions about the determinants of turnover and the role of behavioral factors in future work. The strategy of combining firms’ personnel records with innovative and detailed survey information documented in this chapter indicates another potentially highly fruitful area for future work. While the second chapter exploits detailed information on a single firm to gain insights into particular aspects of the structure and effect of HR practices, sometimes representative surveys also contain information on both workers and firms amenable to a study of certain personnel policies. Such is the case with the French survey of Computerization and Organizational Change that collects information on team production and evaluation strategies for a small sample of workers in a representative sample of French firms. The third chapter, by Diaye, Greenan, and Urdanivia, exploits this source of information to test a model of the effect of subjective evaluation of workers’ effort via interviews in the context of team and group work. Although increasingly common in practice, the analysis of evaluation interviews is rendered difficult due to complex interactions of various incentives, a lack of appropriate data, and identification issues due to unobserved heterogeneity. The chapter extends existing theoretical work to derive various predictions of how evaluation interviews should affect workers’ effort and pay and uses propensity score methods to test this prediction, controlling for selection. The results indicate that evaluation interviews both attract high-productivity workers and have a direct effect on productivity within worker type. As survey data sets incorporate more detailed measures on work effort, organization, and HR practices, similarly ingenuous combinations of specific theoretical modeling and statistical methodologies aimed at identification in other areas of personnel economics should become more common.

6

Stefan Bender et al.

Firm Differences in Human Resource Practices The chapters in this section provide evidence on differences in human resource practices by firms from a variety of different perspectives: over time (von Wachter and Bender), across firms (Hellerstein, Neumark, and McInerney), and within a firm (Manchester). How much firms affect wages has been at the core of the literature analyzing the interrelationship between firms and workers. The evidence is clear that such firm-specific factors as firm size, unionization, and industry structure have important impacts. The interesting contribution of the von Wachter and Bender chapter is to show for a broad sample of establishments that some firm-specific effects are a function of the initial entrylevel conditions that exist when a cohort of workers is hired. As the authors point out, there are two possible reasons for these entry-level differences. The first of these is that the quality of jobs and career opportunities differ for two cohorts; the second is that the degree of rent sharing between workers and firms depends on the extant market conditions at the time of entry. Their analysis exploits their longitudinal information on individuals, together with the firms that hire them, to show that at least some of the substantial wage differences that are observed across firms are due to transitory rents and disappear over time. Another strand of the literature has focused on the role of labor market segregation in explaining race and sex wage differentials. A related strand has focused on workplace segregation by skill, as the productivity of moreeducated workers has increased relative to less-educated ones. However, much of the work has focused on explaining cross-sectional differences across establishments, with the inevitable resulting concerns about omitted confounding factors. The work by Hellerstein, Neumark, and McInerney provides extremely useful initial evidence about both levels of segregation in the United States in 1990 and 2000 as well as changes. They find that racial and ethnic segregation is pervasive. And, while observed segregation by ethnicity has changed little, segregation by race has increased substantially. Most interestingly, the longitudinal nature of their data make it possible for them to show that the increase in racial segregation has been exacerbated by the entry and exit of establishments and by the changing industrial composition of the United States. Hellerstein, Neumark, and McInerney find that segregation by sex, even after controlling for occupational differences and despite countervailing industrial changes, has declined. Segregation by education is also substantial and has increased slightly over the decade. The third chapter advances our understanding about a third HR practice in which substantial firm heterogeneity has been observed: training. Becker’s seminal work suggests that firms will only provide specific train-

Introduction

7

ing and that workers will bear the full cost of general training as general training will increase the likelihood that workers will leave the firm. Despite this theoretical prediction, there is not only abundant evidence that many firms offer tuition reimbursement programs, but also evidence that firms do this to reduce turnover. The Manchester chapter uses primarily a case study approach to examine whether there is an empirical basis that supports the notion that turnover is reduced. She finds that the five-year separation probability of workers who get tuition reimbursement is reduced by over 50 percent, but suggests that the reason is that the investment is complementary to firm-specific human capital. Hence, her results reconcile empirical observation with the Becker theoretical prediction. Effects of Ownership Changes on the Organization of Production While the previous sections featured chapters that examined differences in the ways in which firms treated workers, this section turns to examining what happens within firms as a result of changes in such fundamental features as ownership structure. The transition experienced by economies of Central and Eastern Europe provides a unique opportunity to examine such effects. The two studies on the effects of ownership in this volume, while quite different in their approach to the subject, both represent work that advance our understanding of one such key aspect of transition, the effects of ownership on wages and HR policies. The chapter by Earle and Telegdy uses a large linked employer-employee database for Hungary to analyze the effects of state and foreign ownership on wages. The chapter by Friebel and Panova is a case study of the HR policies in one insiderprivatized Russian firm. A key advantage of the Earle and Telegdy chapter, relative to previous studies of the effects of ownership that are based on either firm-level or worker-level data, is the use of longitudinal employer-employee data, which allows for identification of ownership effects taking into account differences in worker characteristics as well as nonrandom selection of firms into ownership status. The raw estimates show large wage differences across ownership types, but the authors find that ownership type is highly correlated with the education, experience, gender, and occupation of workers, suggesting ownership type may be systematically selected with respect to such characteristics. Nonetheless, the large unconditional wage gaps in the data are little affected by conditioning on worker characteristics. Ownership type is also correlated with firm size, industry, and productivity; controlling for industry reduces the estimated gaps, and controlling for employment size reduces them further. The chapter also exploits the presence of many switches of ownership type in the data to estimate firm fixed effects and random trend models, accounting for unobserved firm characteristics affecting the average level and trend growth of wages. The results

8

Stefan Bender et al.

from these specification differ little in their implications for the effect of privatization, but they reduce the estimated effect of foreign ownership. Overall, the results imply that the substantial unconditional wage differentials are mostly, but not entirely, a function of differences in worker and firm characteristics and that linked panel data are necessary to take these correlated factors into account. Why the operations of firms governed by private owners are quite different from those that are foreign-owned is something that has received a great deal of attention in the literature. While the Hungarian privatization process was not insider-dominated in the same way as other East European economies, most notably Russia, one hypothesis is that insider privatization results in little behavioral change. The fact that insiders benefited much from privatization in Russia raises suspicion about the efficiency of some of the privatization policies, and results suggest that insider-privatized firms do not restructure. However, the study by Friebel and Panova sheds new light on this puzzle in the transition literature by drilling down into the HR practices and the internal labor market of a single insider-privatized Russian firm. The results show career paths prior to transition that are quite similar to the career paths in western firms. In contrast to previous beliefs they find strong micro-evidence for restructuring activities after insider privatization. The employment of blue-collar workers decreased substantially, white-collar workers are recruited from outside the firm, while incumbent white-collar workers are shifted across functions within the firm (but do not leave the firm). As a result, the firm becomes “top loaded,” and career paths from lower levels in the hierarchy are effectively blocked, which consistent with the internal labor market literature can have adverse effects on the efficiency of the firm (see Gibbons and Waldman 1999). In summary, privatization of state-operated enterprises is a key aspect of transition. Both these chapters present evidence that firm governance affects the HR practices and efficiency of firms, but perhaps in more complicated ways than economists thought when first tasked with forming transition policies in Central and Eastern European countries. Indeed, these two chapters highlight the relative unpreparedness of the economics profession to fully understand the impact of transition policies. Globalization, Trade, and Labor Markets Economists have long theorized about the impact of globalization and trade on the earning and employment outcomes of workers. Unfortunately, because the impacts of trade are typically measured at the firm level, and the policy interest is on the long-run outcomes of workers, little data have been available to examine the impact. These three chapters provide some of the first empirical evidence on the topic in analytic work only possible because of the existence of linked longitudinal employer-employee

Introduction

9

(LEE) data.3 The chapters also highlight the new availability of LEE data, which were only available in Northern Europe a decade ago. Muendler uses Brazilian data, the Becker and Muendler chapter examines German multinationals, and Van Biesebroeck uses data from three African nations. The very thought provoking piece by Muendler sets out to examine an issue of concern in every country: the impact of trade liberalization on workers. While economists at the International Monetary Fund (IMF) and the World Bank have long advocated liberalization as a path to economic growth and prosperity, arguing that the resulting reallocation of resources and economic growth would result in the absorption of displaced workers in the growing part of the economy. The rise of socialist leaders throughout South and Central America attest to the lack of popular confidence in such economic theories. Brazil’s experience with trade liberalization in 1990 provides a useful opportunity to examine the facts. Muendler uses LEE data to examine the long-run outcomes of individual workers who worked in firms directly subject to foreign competition (i.e., in sectors in which foreign import penetration increased substantially) and compared them to observationally equivalent individuals in observationally equivalent firms. He finds, as expected, substantial displacement of workers; they also find that neither comparative-advantage sectors nor exporters absorb displaced workers for years. In addition, firms in the new-growth sectors have significantly more displacements and significantly fewer accessions than the exiting firms in the import-competing sector. As a result, workers are much more likely to transition to the informal sector and unemployment. Spells of unemployment last longer, and spells in the formal sector are much more likely to fail. Another hotly debated impact of globalization is the outsourcing of jobs by multinational enterprises (MNEs). Lou Dobbs, a CNN reporter wrote a recent book, Exporting America, which has received enormous attention in the United States. He vehemently argues that too many U.S. companies are sending American jobs overseas and choosing to employ cheap overseas labor—going so far as to list “job exporters” on his Web site. A very different picture is painted in the Becker and Muendler chapter that uses German LEE data to examine the facts in some detail. They find that MNEs that increase their foreign direct investment (FDI) exposure become more competitive, and the resulting expansion acts to significantly reduce the rate of job loss. Indeed, the annual separation rate of workers at MNEs is about 14 percent, compared with the 18 percent separation rate of nonMNEs. One important result that is inconsistent with prior expectations 3. Although there has been some work on this topic by Lori Kletzer (http://econ.ucsc.edu/ Faculty/facLkletz.shtml) or by Chris Ruhm (http://www.nber.org/papers/w5621), the studies are essentially based on worker surveys. Muendler is the first to use LEE data to analyse the long-term effects in a developing country.

10

Stefan Bender et al.

about the nature of Germany’s comparative advantage is that job savings are greater for more-educated than less-educated workers. Although the previous two chapters deal with employment outcomes, policymakers and economists are equally interested in the impact of globalization on earnings. This is particularly interesting in the case of Africa, which has been receiving increasing attention in the development literature, but is probably underresearched due to lack of adequate data. The Van Biesebroeck chapter examines patterns in earnings outcomes of workers in manufacturing plants in three countries that differ substantially in levels of economic development: Zimbabwe, Kenya, and Tanzania. He finds that the more developed a country is (and the more exposed to foreign markets), the more wages match the productivity of the individual worker, controlling for other characteristics. Intriguingly, given the importance placed on education as an investment strategy for developing countries, wages substantially exceed productivity for the most-educated workers in the least-developed countries. This research provides some intriguing evidence that suggests that more-developed countries are also likely to have more efficient labor markets—although obviously the direction of causality is difficult to establish. Data Sources The data sets used in this book are truly international in flavor—and in a sense provide a world tour of the interrelationships between firms and workers. Four chapters (Andersson, Brown, Campbell, Chiang, and Park; Hellerstein, Neumark, and McInerney; Burks, Carpenter, Götte, Monaco, Porter, and Rustichini; and Manchester) use data on U.S. firms and workers. One of the first analyses of South American data is provided by the Muendler chapter, which studies Brazil. Crossing the ocean to Africa, Van Biesebroeck’s chapter provides a study using data from Tanzania, Kenya, and Zimbabwe. The world tour continues with a trip north to the western part of Europe, with chapters from France (Diaye, Greenan, and Urdanivia) and Germany (Becker and Muendler; Bender and von Wachter). It ends in Eastern Europe (Russia, Friebel, and Panova) and Hungary (Earle and Telegdy). One of the themes of the conference was to see how qualitative data analysis could be used to examine employers and employees. There are three case studies in the book; two are based on administrative records of the firm (Manchester; Friebel and Panova). Burks, Carpenter, Götte, Monaco, Porter, and Rustichini is a statistical case study of a single trucker firm and its employees, which matches proprietary personnel and operational data. They combine traditional survey instruments with behavioral economics experiments. The other chapters contribute to the second theme of the book, which

Introduction

11

was quantitative data, primarily LEE data. Although each chapter has the LEE component in common, a major contribution of the book is the description that the authors provide of how they create their LEE data sets by combining different sources with different identification techniques. The basic approach is well illustrated by von Wachter and Bender, who use a single data source for building up their linked employer employee data. And, although most of the chapters use different official data sets and are linked over a unique firm identifier, two chapters (Becker and Muendler; Hellerstein, Neumark, and McInerney) illustrate the use and validity of alphanumerical matching algorithms, such as Automatch, based on the name and addresses of the firms. A different approach is used by Earle and Telegdy who have neither identifiers nor addresses as a basis for linking one data set: they use subsets of common variables to get unique firms per cell for matching. Other striking features are evident upon examination of the chapters. The first is the imaginative set of methods whereby data can be collected from individual firms; the second is how many more data sets have become available since the seminal review by Abowd and Kramarz in 1999; and the third is the number of innovative approaches taken to enhance the breadth and depth of information derived from linked data. Several chapters illustrated the potential richness of looking at single firms in detail. One example is the chapter by Manchester, who analyses the impact of tuition reimbursement programs by examining data from a single firm as well as from a cross section of firms. A panel of observations was constructed based on seven point-in-time observations from administrative records for the firm. Individuals are observed on December 15 of each year from 1999 to 2005, and the data include gender, age, and race as well as start date, job characteristics, and annual wage rates. The chapter by Friebel and Panova also uses the raw data from the HR department of the firm they study. There is one personnel file for each individual in the firm. This file contains information about entry into the firm; exit date (that is, separation); dates of movements across job titles; an occupational code; the departmental affiliation; as well as such personal characteristics as age, education, gender, place of birth, place of university education, field of study, party and trade union membership, ethnicity, marital status, and number of children. Finally, the firm provided them with information about the job history of each individual: military service, date of leaving previous job, last employer. In probably the richest case study analysis in the book, the chapter by Burks, Carpenter, Götte, Monaco, Porter, and Rustichini show the enormous potential of a detailed data collection effort. The team matched proprietary personnel and operational data to new data collected by the researchers to create a two-year panel study of a large subset of new hires that included both standard survey instruments and the results of behavioral economics experiments. The team collected information on employees that

12

Stefan Bender et al.

went beyond the standard demographic information to also capture information on risk and loss aversion, time preference, planning, nonverbal IQ, and the Multidimensional Personality Questionnaire (MPQ) personality profile. The project, which is in its beginning stages, will follow employees over two years of their work lives. Among the major design goals are to discover the extent to which the survey and experimental measures are correlated and whether and how much predictive power, with respect to key onthe-job outcome variables, is added by the behavioral measures. Table I.1 provides a brief summary of the LEE data sets described in each chapter, together with a reference to the Abowd and Kramarz (1999) paper. As is to be expected, data sets that already existed have been considerably updated and enhanced; in a number of cases, new data sets have appeared. The basic richness of LEE data is well illustrated by the von Wachter and Bender chapter: the size and universality of the data mean that they can track the outcomes of quite narrowly defined subgroups of workers as well as the firms that employ them. Their data includes information on all employees covered by social security, representing around 80 percent of the German workforce, with detailed histories for each worker’s time in covered employment. Their data are unusually rich for LEE data as they not only include basic demographic information, as well as data on occupation, industry, job status, education, and individual-level information on gross daily wages subject to social security contributions and the exact dates when the employee worked in the social security system. The unique establishment identifiers available were used to create a separate data set of establishment characteristics that were aggregated up from the employment register and merged back onto the individual-level data. Characteristics include, among others, establishment size, employment growth, and average wages. Muendler’s chapter features one of the first LEE data sets for a developing country: Brazil. The data set is derived from a nationwide, comprehensive set of administrative records of workers employed in the formal sector. The ministry of labor estimates that well above 90 percent of all formally employed workers in Brazil are covered in Relação Anual de Informatições Sociais (RAIS) throughout the 1990s. As is typical with LEE data derived from administrative records, the sample size is enormous: information on 71.1 million workers (with 556.3 million job spells) at 5.52 million establishments in 3.75 million firms over the sixteen-year period 1986 to 2001. The data also have the rich feature of providing the month of accession and the month of separation in addition to such other relevant worker information includes tenure at the establishment, age, gender, and educational attainment; job information includes occupation and the monthly average wage; establishment information includes sector and municipality classifications. As is also often the case with LEE data derived from administra-

Introduction Table I.1

Author

Summary of linked employer-employee data sets

Country

Brief description

Andersson et al.

United States

Diaye et al. Von Wachter & Bender Hellerstein et al.

France Germany United States

a. Longitudinal individual data from the LEHD b. 1997 Economic Censuses c. The 1991–1998 Census/NSF R&D Surveys (firm-level R&D) COI survey Social security records—total population 1990 and 2000 DEEDs

Earle & Telegdy

Hungary

a. Hungarian Wage Survey

Muendler

Brazil

b. Tax Authority data RAIS

Becker & Muendler

Germany

Van Biesebroeck

13

Tanzania, Kenya, and Zimbabwe

a. Social security records of the German Federal Labor Agency b. Midi database (MIcro database Direct Investment, formerly direk), collected by Deutsche Bundesbank c. Commercial corporate structure database markus (from Verband der Vereine Creditreform) A stratified sample of manufacturing firms in three consecutive years

Abowd & Kramarz comparison New data set

Enhanced data set Referenced in Abowd & Kramarz Updated and expanded New data set, new country New data, new country Updated and enhanced (by b and c)

New data sets, new countries

tive records, the basic data set can be enhanced by matching into other administrative records—in this case, export data. The Andersson, Brown, Campbell, Chiang, and Park chapter further illustrates the potential for enhancing basic LEE data by using multiple data sources. In order to answer their research question, they use data from three sources: longitudinal and near-universal individual data from the Longitudinal Employer-Household Dynamics Program (LEHD) program to construct and characterize the HR practices of firms; they add firm characteristics from the 1997 Economic Censuses (e.g., measures of revenue, material costs, total hours, capital stock, four-digit industry code) as well as from the 1991 to 1998 Census National Science Foundation (NSF) R&D Surveys (firm-level R&D). A very similar approach is taken by Becker and Muendler, who construct their LEE set from three confidential micro-data sources, from multiple sources, and complement them with sector-level information on

14

Stefan Bender et al.

German foreign trade. This chapter illustrates some of the matching challenges common to LEE data, in that the three data sources do not share common firm identifiers. The authors surmount the problem by using a string-matching procedure to identify clearly identical firms and their establishments. The result of their efforts is to create an impressively large data set that constitutes a cross section of establishments around year 2000, including a total of 39,681 establishments whose German parent firms conduct FDI abroad and 1,133,920 control establishments out of 3.8 million establishments in the full worker sample (1998 to 2002). Not all LEE data are derived from official administrative records. Indeed, Van Biesebroeck’s uses stratified samples of manufacturing firms in three African countries that were collected by three different research teams, coordinated by the Regional Program of Enterprise Development at the World Bank. When working with survey data, one of the most difficult decisions is whether to choose a sample that is representative of firms or one that is representative of workers. In this case, the latter was chosen: firms were sampled to give (the firm of) each manufacturing worker equal probability to be included in the sample—an implicit stratification by employment size. The second choice is how many employees to interview within each firm. The cost constraints that are an integral part of survey work impose a trade-off between firm and worker diversity: the more employees interviewed in a firm (and, hence, the more within firm diversity is captured), the fewer firms can be interviewed (resulting in less across-firm diversity). In the African countries, a maximum of ten employees per firm were interviewed each year, resulting in an unbalanced panel of firms with, on average, 110 to 183 observations per year in each country. Another example of a survey approach is evident in the Diaye, Greenan, and Urdanivia chapter, which analyzes a French matched employeremployee survey on computerization and organizational change. In contrast to the African case, however, the sampling unit is the firm, and the frame is a representative sample of manufacturing firms with more than fifty employees and a sample of randomly selected employees within these firms. The French sample chose the opposite trade-off from the African, interviewing a small sample of employees (one, two, or three) within each firm and, hence, getting more firm diversity. An interesting feature of the French survey, and one that emphasizes the value added of surveys relative to administrative records, is that the labor force section provides a detailed description about the organization of work, particularly whether that work is structured around group activities. In addition, the survey captures different measures of effort, which would be impossible to capture in administrative records. The Earle and Telegdy chapter also draws on data from multiple sources. The worker data come from the Hungarian Wage Survey, compiled by the National Employment Office and maintained by the Institute of Econom-

Introduction

15

ics of the Hungarian Academy of Sciences. The latter organization also links these data to some firm-level information. The authors have further linked this database (using common variables in both databases) to a universal tax database containing detailed information on all Hungarian firms using double-sided accounting. Employers are included in the Wage Survey according to whether their employees are selected by a random procedure: in the first two years of the survey, workers are selected using a fixed interval of selection, while subsequently workers are selected by birthdate. Firms were included only if they have employees born on these dates. Although this approach provides a random sample of workers within firms and includes, on average, about 6.5 percent of production workers and 10 percent of nonproduction workers, the sample of firms is related to size. The authors, therefore, use weights related to size and response probabilities in their analysis, and the final sample consists of a panel of 21,238 firms linked with a within-firm random sample of 1.35 million worker observations. The final approach that can be used to create LEE data sets is to match existing surveys with existing administrative data. This approach is described in the Hellerstein, Neumark, and McInerney chapter. They draw a sample of workers from the Sample Edited Detail File (SEDF), which contains all individual responses to the Decennial Census of Population onein-six Long Form. Respondents are asked to provide the name and address of their employer in the previous week. This information is then matched to the name and address information in the Census Bureau’s Business Register list (BR), based on administrative records. This procedure yields a very large data set with workers matched to their establishments, along with all of the information on workers from the SEDF. Even more interesting, because the data are constructed for both 1990 and 2000, the authors are able to examine changes in establishments over time. Summary and Outlook In summary, the chapters in this volume all represent research that relies on advances in data collection methods in one way or another. These range from combining case study data with personnel records of a single firm, ideally suited to understand issues such as how HR policies affect workers and the performance of firms, to the creation of new multi-source, nationally representative LEE data sets, ideally suited to capture empirical irregularities related to the dynamics of the economy. It is easy to envision further advances along each of these lines, but perhaps even more promising is the crossing of the two paths. The studies in this volume clearly demonstrate the relative advantages of the two approaches and the results often complement each other in a way that adds to our overall understanding. The benefits of being able to combine data from the two study approaches within a single framework are obvious, as is the increasing po-

16

Stefan Bender et al.

tential for cross-national comparisons (see Hamermesh 2007). Today they are more feasible than before thanks to new data collection tools and new integration techniques. Indeed, active research is in progress to match data sets from many different sources to inform research about active labor market policies, price setting, and employment dynamics. Similarly, international comparisons of LEE and firm surveys are beginning to emerge (Lazear and Shaw, forthcoming; Freeman, Kruse, and Blasi, forthcoming). However, there are clearly important and big challenges in terms of data access issues and disclosure avoidance that need to be addressed. The creation and analysis of high-quality information are core elements of the scientific endeavor. No less fundamental is the ability to replicate scientific analysis. Yet the firm-level data that is described in this book are often not accessible to others for replication and validation. It will be critical to develop widely available access modalities for the qualitative data, often housed at universities and research centers, and administrative data, often housed at statistical agencies. Only recently has the research community begun to address such key issues that will ultimately affect the scientific nature of the research as well as our ability to access and gather new data. Some progress has already been made, including the development of remote access capabilities and new synthetic data methods, but this is likely to remain a key challenge for the research community in the foreseeable future.

References Abowd, John, John Haltiwanger, and Julia Lane. 2004. Integrated longitudinal employer-employee data for the United States. American Economic Review 94:224–29. Abowd, John, and Francis Kramarz. 1999. The analysis of labor markets using matched employer-employee data. In Handbook of labor economics. Vol. 3B, ed. O. Ashenfelter and D. Card, 2629–2710. Amsterdam: North-Holland. Brown, Clair, John Haltiwanger, and Julia Lane. 2006. Economic turbulence: Is volatility good for America? Chicago: University of Chicago Press. Freeman, Richard B., Douglas Kruse, and Joseph Blasi. Forthcoming. Worker reports on labour practices and outcomes in a single form across countries. Labour Economics. Gibbons, Robert, and Michael Waldman. 1999. A theory of wage and promotion dynamics in inside firms. Quarterly Journal of Economics 114 (1999): 132–58. Hamermesh, Dan. 2007. Fun with matched employer-employee data. Labour Economics, forthcoming. Lazear, Edward P., and Kathryn L. Shaw, eds. Forthcoming. An international comparison of the structure of wages. Chicago: University of Chicago Press.

1 The Effect of HRM Practices and R&D Investment on Worker Productivity Fredrik Andersson, Clair Brown, Benjamin Campbell, Hyowook Chiang, and Yooki Park

1.1 Introduction As the pace of technological change has quickened and global competition has shortened product life cycles, firms have had to rethink their technology investment strategies and their human resource management practices in order to remain competitive. The main contribution of this chapter is to examine the relationship between firm-level technological advanceFredrik Andersson is a senior research associate of the Cornell Institute for Social and Economic Research, and a research fellow of the Longitudinal Employer-Household Dynamics Program (LEHD) of the U.S. Bureau of the Census. Clair Brown is a professor of economics and director of the Center for Work, Technology, and Society (IIR) at the University of California, Berkeley. Benjamin Campbell is an assistant professor of management and human resources at the Fisher College of Business, Ohio State University. Hyowook Chiang is an economist at Welch Consulting. Yooki Park is an associate at McKinsey & Company. The data used are confidential data from the U.S. Census Bureau’s Longitudinal EmployerHousehold Dynamics Program (LEHD), which is partially supported by the National Science Foundation Grant SES-9978093 to Cornell University (Cornell Institute for Social and Economic Research), the National Institute on Aging (R01-AG18854-01), and the Alfred P. Sloan Foundation. Support was also provided by the Institute of Industrial Relations at the University of California at Berkeley, and the Institute for Technology, Enterprise, and Competitiveness (ITEC/COE) and Omron Fellowship at Doshisha University. We have benefited from discussions with and comments from Charlie Brown, Peter Cappelli, Erica Groshen, Andrew Hildreth, Julia Lane, Daniel Parent, Linda Sattler, Eric Verhoogen, Till von Wachter, and Edward Wolff, seminar participants at Berkeley and Wharton, participants at the NBER Summer Institute and the Conference on the Analysis of Firms and Employees (CAFE), and the anonymous reviewers. This document has undergone a Census Bureau review more limited in scope than that given to official Census Bureau publications and is released to inform interested parties of ongoing research and to encourage discussion of work in progress. The views expressed herein are attributable only to the author(s) and do not represent the views of the U.S. Census Bureau, its program sponsors, or data providers. The U.S. Census Bureau is preparing to support external researchers’ use of these data; please contact U.S. Census Bureau, LEHD Program, FB 2138-3, 4700 Silver Hill Rd., Suitland, MD 20233, USA.

19

20

Fredrik Andersson et al.

ment (as proxied by research and development investment [R&D]) and firms’ human resource management (HRM) practices for high-skill workers in a high-tech industry, and then examine how this relationship is connected to firm performance. Although the relationship of technological change and labor market outcomes at the individual-worker level has been well-studied,1 surprisingly little is known about what happens within the firm. Specifically, there is little empirical research on whether firms’ technology choices are consistent with their human resource practices and whether there is a statistical relationship between technology, human resources, and performance at the firm level. At the individual level, there is a long line of research observing the correlation of technical change and compensation for high-skill workers and examining the mechanisms underlying the relationship.2 However, there is little large-scale work looking at the relationship of technology and worker outcomes within firms. In this project, we examine the worker/technology relationship within firms and focus on one specific industry where we can employ detailed industry controls. Previous research has demonstrated that technology interacts with human resource practices through several channels. Technology may alter the development of and returns to human capital (Krueger 1993; Handel 1999; DiNardo and Pischke 1997; and Entorf and Kramarz 1998). Additionally, technology can interact with individual outcomes through changes in work design (Hunter and Lafkas 2003; Bresnahan, Brynjolfsson, and Hitt 2002; Zuboff 1988; Autor, Levy, and Murnane 2002; Brown et al. 1997; and Barley and Orr 1997) or changes in work organization (Cappelli 1996; Bresnahan, Brynjolfsson, and Hitt 2002; O’Shaughnessy, Levine, and Cappelli 2001; and Caroli and Van Reenen 2001). We propose a mechanism connecting technology and HRM practices at the firm level that links the skill bias and the organization change approaches. We propose a make-versus-buy model of workforce skill adjustment. If technology and labor force skills are complements in a firm’s production function, and if the HRM system impacts the cost of acquiring, developing, and retaining the portfolio of skills in a firm, then the firm’s choice of HRM system affects its ability to adjust worker skill levels to maximize the value of their technological investments. In other words, if firms choose to augment the skill of their workforce to complement an investment in technology, they face the traditional make-versus-buy problem. Firms can structure their HRM practices to develop and retain the 1. See Brown and Campbell (2002) for a detailed review of the impact of technological change on the work and wages of individuals. 2. Seminal works in this area include Bound and Johnson (1992), Levy and Murnane (1992), Katz and Murphy (1992), Juhn, Murphy, and Pierce (1993), Berman, Bound, and Griliches (1994), and Allen (1997).

The Effect of HRM Practices and R&D Investment on Worker Productivity

21

necessary skills in-house, or they can structure their HRM practices to attract and recruit workers with the necessary skills on the external market. For the econometric portion of the chapter, we utilize data from the Longitudinal Employer-Household Dynamics (LEHD) program that covers all establishments and their employees in seven large states over the period 1992 to 2001. Our analysis and interpretation is informed by fieldwork through the Sloan Competitive Semiconductor Manufacturing Program. The fieldwork began in 1992 and involved detailed data collection and intensive interviews at over three dozen semiconductor manufacturing firms. The insights and understanding developed through these site visits are the basis of our examination of the HRM-technology-productivity relationship in manufacturing firms within the electronics industry. The analysis capitalizes on the strengths of both detailed industry study and large-scale survey approaches to develop a framework for estimating the relationship between firm productivity, R&D, and HRM practices using near-universal data from the LEHD program. The detailed industry knowledge facilitates interpreting of the results and understanding the context in which the results are embedded. This combined method expands the “insider econometrics” approach to the industry level of analysis. Industries exhibit wide variation in their market and organizational structures, which affects their rate of technological change, their degree of and response to global competition, and, in general, their reaction to environmental factors. Because industries differ in dimensions that are hard to measure using traditional large-scale survey responses, industry-specific knowledge gained through fieldwork is critical in understanding how to interpret estimated statistical relationships within and across industries (Brown, Haltiwanger, and Lane 2006). Fieldwork research allows us to study the trade-offs that managers make in developing and implementing new technology and HRM practices, and they help us understand the timing of these decisions. Although technology and HRM practices are related through the production process, we observe that technological changes can be implemented much faster and, therefore, more often than HRM practices can be changed. Firms’ choice of HRM practices can be made more independently of the external market than the choice of a new technology, which is tied to a choice of customers and markets. While it is important to understand the underlying structure of a firm to interpret results, detailed firm or industry studies cannot be used for generalizations across firms or industries or for estimating national impact because we do not know whether the specific firm experiences are representative (Sturgeon et al. 2006). In order to generalize from an industry study, we need estimations from a comprehensive survey across firms and workers in the industry that show the observed relationship across key variables, such as HRM practices, R&D expenditures, and productivity. Our approach is to combine comprehensive micro-data and detailed industry

22

Fredrik Andersson et al.

knowledge in a way that leverages each approach’s strengths and increases the quality and usefulness of both types of data. For micro-data to be estimated and interpreted properly, research teams must have both a deep understanding of the industries involved and expertise in the use of microdata. Specifically, this study combines both micro-data and detailed industry knowledge to analyze the impact of R&D and HRM systems on firm performance within the electronics industry (Standard Industrial Classification [SIC] 35 and 36).3 Although firms in the electronics industry have a high level of R&D investment relative to other industries, there is a large variance in investment between firms within the industry. Studying one industry simplifies the analysis of the relationship of R&D and HRM by focusing on firms that are fairly comparable in structure and face similar market trends and measurement issues. We use workers’ outcomes within establishments to construct a variety of measures of establishment-level HRM outcomes for high-education workers and then link these HRM measures to plant and firm characteristics. First, we document the firms’ HRM systems. Implementation of HRM systems is more important than implementation of individual components because there are synergies and complementarities in HRM practices (Kandel and Lazear 1992; and Milgrom and Roberts 1995). We perform a cluster analysis of firms and HRM measures to identify and describe the most common HRM systems. Next, we employ principal components analysis to identify groups of correlated HRM measures. We then regress worker productivity on the principal HRM components interacted with R&D. We find substantial variation in HRM practices for high-education workers across firms in this industry. Human resource management bundles appear to include both spot market and internal labor market outcomes. Consistent with Bauer and Bender’s (2004) finding using comparable German data that technological advancement is correlated with worker churning for high-skilled workers, we find that for firms with high levels of R&D, HRM practices that provide multiple ports of entry, low turnover and performance incentives are positively related to worker productivity. For low R&D firms, HRM practices that offer multiple ports of entry and low wage variance for recent hires are positively related to productivity. Additionally, the results indicate strong clustering of HRM practices across firms, with high R&D firms much more likely to implement more market-oriented practices than low R&D firms in this high-tech industry. These findings are consistent with the implications of our make-versus-buy model of workforce skills, where firms with a high rate of technological change that buy 3. The LEHD program links universal and longitudinal records on employees’ earnings and employment from states’ Unemployment Insurance (UI) systems with detailed crosssectional data from Census Bureau’s Economic Censuses and Census/NSF R&D surveys.

The Effect of HRM Practices and R&D Investment on Worker Productivity

23

new skills on the external market and selectively retain and reward experienced workers will demonstrate higher productivity than comparable firms with fewer ports of entry with similar earnings growth, which indicates more internal skill development. Also, firms with a low rate of technological change implement HRM systems that are consistent with providing training to workers rather than buying required skills, which may be more efficient in these firms with slower technological change. The next section presents a framework for firms’ R&D investment decisions and firms’ HRM decisions and how these decisions are related to productivity. Then we describe the data set and our measurements for HRM practices, R&D investment, firm performance, and other firm characteristics. We present statistical results on firm performance, HRM, and R&D and discuss to what extent the results are consistent with our hypotheses. Finally, we conclude with a summary and a discussion of the implications of the research. 1.2 HRM Practices and Workforce Skill Adjustment Costs Our analysis looks at HRM practices within firms and builds on the Internal Labor Market analysis embedded in the work of Prendergast (1996) and Doeringer and Piore (1971). In the empirical work, there is mixed evidence on measuring internal labor markets within firms. Using data from a single firm, Baker, Gibbs, and Holmstrom (1994) find that some aspects of the employment relationship are consistent with the theory of internal labor markets. Lazear and Oyer (2004) use matched data from the Swedish Employers Confederation from 1970 to 1990. They find that the strict model of internal labor markets does not seem to hold because external forces play a large role in firms’ wage setting policies. Topel and Ward (1992) observe high mobility and earnings growth among young male workers that is more consistent with matching models and on-the-job search than internal labor markets. Because of the mixed evidence, we perform a cluster analysis of firms in our sample to examine the distribution of different sets of HRM practices and find a diverse set of HRM outcomes, even within a homogenous industry. Given the diverse outcomes, we focus on developing an understanding of the underlying process that might explain the diversity. The basic concept of the framework is that HRM practices affect the cost structure of how firms adjust the skills of their workforce. If technology and worker skills are complementary, then the firm’s HRM decisions and R&D decisions will be related. Even in the high-tech electronics sector, the speed of technological change varies across firms in different product markets. For example, consider the semiconductor industry, which is one of the industries included in our sample. Within the semiconductor industry, graphic chips for video games typically have a generation life of approximately eighteen months

24

Fredrik Andersson et al.

and analogue chips typically have a generation life of five years. Memory chips and microprocessors typically have a generation life between two and three years. Generation life is critical in defining a firm’s constraints in making technological investment, as product prices are above marginal costs early in the cycle before supply brings the prices down. Across the electronics industry more broadly, product life and speed of technological change have an even longer time horizon. For example, our sample also includes manufactures of “current-carrying wiring devices.” In contrast to the semiconductor industry, the wire industry is marked by very long product life spans and low levels of innovation. The firm’s HRM system structures how labor inputs are bought and created over time. We assume the cost of labor inputs are determined by the following HRM practices:

• • • •

Screening and hiring Skill development (both learning by doing and formal training) Retention of experienced workers Adjustments in headcount by skill (quits and layoffs)

At any given point in time, these HRM practices determine the cost and skills of the firm’s workforce. Here we focus only on high-education workers because they are the workers who develop and implement new technology. If firms adopt a technological change that alters the optimal composition of their workforce, firms may choose to adjust the skills embedded in their workforce. Given the decision to adjust workforce skills, firms must make two major decisions in creating the optimal skill-experience composition in the workforce: 1. Decide whether to provide formal training in the new technology to their existing workers or to purchase these skills through new hires (we call this the make-buy decision) 2. Decide which experienced engineers (and other workers) they will retain (we call this the retention decision) The firm will make the first decision based upon the relative costs, including both the payroll costs and the time-to-market costs, of making or buying the required skills for the new technology. Under the assumption that the cost of “making” the required skills is the worker adjustment cost of acquiring skills (training cost) and is proportional to the size of technological jumps over a given time, and that the cost of “buying” the required skills is the firm’s adjustment costs in hiring new workers, which is invariant to the size of the technological jump, then for sufficiently large technological jumps, “buying” will be relatively less costly than “making” new skills. The second decision will depend upon the costs of retention as well as the production function. Specifically, firms will structure incentive systems to retain the workers who are most valuable to the firm. For a new tech-

The Effect of HRM Practices and R&D Investment on Worker Productivity

25

nology that requires new skills and restructures skill demand in the firm, the firm must decide which workers to retain. This decision depends on the portfolio of skills supplied in the firm compared to the portfolio of skills necessary for the new technology and the costs of obtaining the new portfolio, which include a comparison of the make decisions (primarily retraining costs) compared to buy decision (cost of new hires, layoffs, and worker morale). The costs to workers of retraining depend on their opportunity wage and the required effort associated with retraining, which depends on how much retraining is required. Workers with skill sets far behind the latest technology will face higher retraining costs but require lower incentives by the firm for retention, while workers who are better matches to the new technology will face lower retraining costs and the incentives required by the firm for retention are higher. How does the firm’s product life, and thus rate of R&D spending, affect how the HRM system operates? We assume that a new technology requires a mix of experience on the previous generation of technology and new skills that require formal education (or training). Firms in short productlife markets, and thus with high R&D spending, must have a mix of engineers with the new skills required for the new technology and engineers with experience on the last generation of technology, and we assume that experience and new skills are complements. Firms in long product-life markets, and thus with low R&D spending, rely more on engineers with experience because the engineers will focus on cutting costs, improving quality, and improving throughput over the life of the product. If worker costs of retraining increases proportionally with size of technological change (as proxied by R&D), and firm hiring transaction costs are invariant to size of technological change, then R&D and flexible hiring practices will be positively related to worker productivity. In a competitive labor market, implementation of new technologies in an industry will impact the external market opportunities for engineers. To counteract turnover of key workers, who are the workers with skills more compatible with the new technology, firms will structure their HRM system to provide incentives (both in compensation and in job assignment) in order to retain workers who match well to the new technology and who face lower personal retraining costs. How long it is beneficial for the high R&D firm to retain and use their technical workers’ skills will determine the incentive structure implicit in their pay system compared to opportunity market wages as well as their average tenure (and turnover rate). Hypothesis 1: Firms with high R&D that choose an HRM system that allows hiring of workers with required skills and fosters retention of selected experienced workers will have higher worker productivity than those that create the required new skills strictly through retraining of workers or strictly through new hires.

26

Fredrik Andersson et al.

Firms with low R&D improve performance not through product market innovation, but through incremental improvement in the product and production process. Experience is valuable in making these improvements, and firms that provide incentives to retain workers will have higher productivity. Performance-based pay that is tied to improvements may also motivate workers to higher productivity, although this pay may be awarded to a team rather than an individual technical worker in order to encourage group activity and because evaluating individual contributions may be difficult. Hypothesis 2: Firms with low R&D that choose an HRM system that fosters retention of experienced workers and allows some performance-based pay will have higher worker productivity than those that do not have a compensation structure that reduces quits and rewards improvements. In the next section, we discuss the data and measures we will use to examine the previous hypotheses linking HRM practices to worker productivity for firms on different technology paths. 1.3 Data Set and Measures We use data from three sources in our analysis. We use longitudinal and near-universal individual data from the LEHD program to construct and characterize the human resource practices of firms; we add firm characteristics from the 1997 Economic Censuses (e.g., measures of revenue, material costs, total hours, capital stock, four-digit industry code) as well as from the 1991 to 1998 Census/National Science Foundation (NSF) R&D Surveys (firm-level R&D). The LEHD data have been extensively described elsewhere (see Haltiwanger, Lane, and Spletzer 2000; Abowd, Haltiwanger, and Lane 2004),4 but it is worth noting that these data have several advantages over household-based survey data. In particular, the earnings are quite accurately reported as there are financial penalties for misreporting. The data are current, and the data set is extremely large. 1.3.1 HRM Variables In characterizing the human resource practices of a firm, we utilize the measures of earnings, earnings growth, accession rate, and separation rate for selected cohorts within each firm to create the following components of 4. The LEHD database consists of quarterly records of the employment and earnings of almost all individuals from the unemployment insurance systems of a number of U.S. states in the 1990s. The UI records have also been matched to internal administrative records at the Census Bureau that contain information on date of birth, place of birth, race, and sex for all workers.

The Effect of HRM Practices and R&D Investment on Worker Productivity

27

firms’ HRM systems for high-education (or professional) workers, who we know from our fieldwork are primarily technical workers:

• Accession rate: Ratio of the total number of new hires to the total number of workers in 1997.

• Ratio of mean initial wage to market initial wage: Average wage of new

• • • • • • •

hires of an individual establishment divided by average wage of new hires of all establishments in electronics industry (SIC 35 and 36) in 1997. Standard deviation of initial earnings: Standard deviation of earnings of new hires in 1997. Separation rate for workers with two years experience: Proportion of workers who are no longer working for a certain establishment in 1997 among all workers who are hired in 1995 at the same establishment. Within-job wage growth for workers with two years experience: Wage growth between 1995 and 1997 of workers hired in 1995. Standard deviation of within-job wage growth for workers with two years experience: Standard deviation of wage growth between 1995 and 1997 of workers hired in 1995. Separation rate of workers with five years experience: Proportion of workers who are no longer working for a certain establishment in 1997 among all workers who are hired in 1992 at the same establishment. Within-job wage growth for workers with five years experience: Wage growth between 1992 and 1997 of workers hired in 1992. Standard deviation of within-job wage growth for workers with five years experience: Standard deviation of wage growth between 1992 and 1997 of workers hired in 1992.

One limitation of the data is that the current observed HRM practices in a firm reflect outcomes for workers who are both new to the firm and have been at the firm for any number of years. To capture the entire profile of workers and their wage growth, it is necessary to use the longitudinal variation in the data in order to construct the HRM measures. Currently the limited data on R&D expenditures allows us to examine only one cross section of the data, while the HRM measures capture longitudinal variation. Another limitation for this study is that we lack direct measures of some important worker and job characteristics, especially education and occupation. We use imputed education values developed by the LEHD staff to distinguish high-education workers from other types of workers.5 5. While data on education for the individuals in our sample are not directly observed, LEHD staff has imputed education for every individual based on probabilistic links to external data. The statistical model takes advantage of the common observable characteristics in LEHD and Decennial data—most important earnings, industry, geography, gender, and

28

Fredrik Andersson et al.

1.3.2 R&D Measure The following variable represents firm-level technology practices: R&D spending rate is measured as the average total R&D costs per payroll over 1991–1998. Because Census/NSF R&D surveys are conducted at the firm level, we assume that all establishments of the same firm equally benefit from their firm level R&D. Research and development is just one component of firms’ technology investment decisions, and as a result it is an imperfect proxy for investment in technology. However, R&D may be a good proxy for picking up firm’s ability to learn and develop new knowledge (Cohen and Levinthal 1989). Also, because the relationship between R&D and new technology depends on the success of the investments and the length of period until implementation takes place, there may be an issue with the timing of investments and HRM choices. We partition the firms in our sample into two sets: firms with above-mean R&D investment and firms with below-mean investment. 1.3.3 Firm Performance Measure To represent firm performance, we use following productivity measurement: Labor productivity is the log of real value added per total hours worked where the value added is the establishment-level revenue adjusted for inventory change net of materials input, and total hours worked include both production worker hours and nonproduction worker hours. In the next section, we identify common HRM systems, the underlying HRM components that differentiate firms’ HRM systems, and the relationship of these components to worker productivity. 1.4 Empirical Analysis First, we perform a cluster analysis of firm HRM practices to identify the most common HRM systems for high-education workers. Next, we employ principal components analysis to identify groups of correlated HRM measures. We then implement a principal components regression to examine the statistical relationship of worker productivity with HRM practices for different technology paths. 1.4.1 HRM Cluster Descriptions Firms implement HRM practices in bundles, and so we expect a highlevel of correlation of adopted bundles across firms. We perform cluster age—to impute education based on draws from the conditional distribution of educational categories in the 1990 Decennial Census. Details of the statistical model can be obtained from the authors.

The Effect of HRM Practices and R&D Investment on Worker Productivity

29

analysis to identify the most common bundles of HRM practices implemented by firms and to group firms with similar practices. In order to maximize the degree of separations between the groups of firms, clusters of firms are based on canonical variables of HRM variables using Ward’s minimum variance method.6 In table 1.1, we present the cluster results for the HRM variables for high-education workers (summary statistics of the first four clusters of HRM practices are reported, and the last group of firms represents the aggregation of multiple small clusters that are not disclosable according to Census Bureau confidentiality requirements). Each cluster represents a prototype HRM system, which we indicate by name. We then discuss how we think the HRM system is operating within the firm based upon the components and our fieldwork observations. • Cluster 1  Performance-Based Internal Labor Market (ILM): Firms in Cluster 1 offer lower than average initial earnings and slow, but steady earnings growth, lower than average turnover and low earnings dispersion. These characteristics are consistent with hiring lessexperienced workers and advancing them along well-defined pay scales. Entry of workers and their initial earnings reflect skill requirements, so average initial earnings of new hires are higher and have higher variance than in a bureaucratic ILM. After approximately two years, workers are selected (based upon performance) for faster career development, and members of a cohort compete for entry into these favored positions, which have higher earnings growth and lower separation rates. Those who do not receive skill development have lower earnings growth and higher separation rates. • Cluster 2  Spot Market with Rewards: Firms in Cluster 2 exhibit pay tied to the external labor market (both at entry and with experience) and above-average turnover. Firms can identify workers’ talents and skills and hire and pay accordingly (matching is good). Firms can monitor worker performance and pay workers according to contribution. Initial earnings and earnings growth reflect market rates for skill 6. In Ward’s minimum-variance method, the distance between two clusters is the analysis of variance (ANOVA) sum of squares between the two clusters added up over all the variables. At each generation, the within-cluster sum of squares is minimized over all partitions obtainable by merging two clusters from the previous generation (Ward 1963). The assumptions under which Ward’s method joins clusters to maximize the likelihood at each level of the hierarchy are multivariate normal mixture, equal spherical covariance matrices, and equal sampling probabilities. Therefore, we first obtain approximate estimates of the pooled withincluster covariance matrix of the HRM variables when the clusters are assumed to be multivariate normal with spherical covariance using the approximate covariance estimation for clustering developed by Art, Gnanadesikan, and Kettenring (1982). The Approximate Covariance Estimation for CLUStering (ACECLUS) procedure provides us with canonical versions of earnings (or person and firm effect), earnings growth, and worker churning that we use in the cluster analysis.

120

273

235

0.131 (0.103) 0.539 (0.265) 2,754 (987) 0.406 (0.197) 0.071 (0.061) 0.116 (0.074) 0.403 (0.197) 0.026 (0.029) 0.053 (0.030)

Cluster 3: Bureaucratic ILM

57

0.135 (0.079) 1.153 (0.180) 13,672 (1,188) 0.486 (0.205) 0.056 (0.076) 0.156 (0.073) 0.446 (0.163) 0.028 (0.025) 0.055 (0.022)

Cluster 4: Spot market

56

0.169 (0.102) 1.437 (0.222) 21,430 (594) 0.435 (0.174) 0.067 (0.067) 0.127 (0.093) 0.531 (0.171) 0.033 (0.031) 0.062 (0.025)

Residual firms

741

0.140 (0.095) 0.830 (0.350) 7,419 (5,939) 0.430 (0.195) 0.060 (0.068) 0.120 (0.076) 0.430 (0.177) 0.030 (0.027) 0.060 (0.024)

Sample

Notes: Table shows within-cluster means. Numbers in parentheses are standard deviations. HRM = human resource management. ILM = internal labor market. SD = standard deviation.

N

SD of within job wage growth at 5 years tenure

Within job wage growth at 5 years tenure

Separation rate at 5 years tenure

SD of within job wage growth at 2 years tenure

Within job wage growth at 2 years tenure

Separation rate at 2 years tenure

SD of initial earnings

0.141 (0.096) 1.027 (0.214) 10,024 (1,029) 0.462 (0.185) 0.066 (0.061) 0.129 (0.076) 0.452 (0.172) 0.030 (0.027) 0.060 (0.019)

Cluster 2: Spot market with rewards

0.141 (0.116) 0.807 (0.382) 6,108 (10,210) 0.414 (0.194) 0.052 (0.071) 0.121 (0.060) 0.425 (0.176) 0.030 (0.029) 0.054 (0.030)

Cluster 1: Performance-based ILM

HRM practice clusters for high-education workers

Ratio of mean initial wage to market initial wage

Accession rate

Variable

Table 1.1

The Effect of HRM Practices and R&D Investment on Worker Productivity

31

and talent, with large initial variance, and variance does not increase over tenure. Separation rate is higher than in ILMs. • Cluster 3  Bureaucratic ILM: Firms in Cluster 3 have very low initial earnings (with low variance), below-average earnings growth the first two years, and low turnover. Initial earnings of new hires are similar as most workers enter at the same level and have similar (and reliable) earnings growth. Firms experience low separation rates. • Cluster 4  Spot Market: Firms in cluster 4 offer high initial wages with large variance, below-average earnings growth (with large variance), and high turnover. Firms hire and pay workers as in spot market, but identification of worker’s talents and effort at hire is imperfect and monitoring of worker performance is imperfect. Variance of initial earnings is higher than in spot markets with rewards. Early separation rate is higher than in spot market because the bad matches (both at hire and in rewards) end. We find that firms are concentrated in clusters 1 and 3: 37 percent of all firms are in Cluster 1, 32 percent are in Cluster 3, 16 percent are in Cluster 2, and 8 percent are in Cluster 4; apparently the primary variables in differentiating systems are wage variation and initial earnings. We classify firms as high or low R&D firms based on whether their R&D investment is above or below the mean and present their HRM cluster distributions in table 1.2. For low R&D firms, 82 percent are in Clusters 1 and 3 (ILMs), and 14 percent are in Clusters 2 and 4 (spot markets); for high R&D firms, 60 percent are in Clusters 1 and 3, and 31 percent in Cluster 2 and 4. This indicates that high R&D firms are more likely to implement spot-market-oriented HRM practices than low R&D firms although there are still many high R&D firms in the performance-based ILM cluster. Further, low R&D firms are more likely than high R&D firms to implement Bureaucratic ILMs, although there are many low R&D firms that implement performance-based ILM.

Table 1.2

High-education HRM cluster sizes by firm R&D level Low R&D firms

Cluster 1: Performance-based ILM Cluster 2: Spot market with rewards Cluster 3: Bureaucratic ILM Cluster 4: Spot market Residual firms

120 34 125 8 13

(16.2) (4.6) (16.9) (1.1) (1.8)

High R&D firms 153 86 110 49 43

(20.6) (11.6) (14.8) (6.6) (5.8)

Notes: Numbers in parentheses are percentages. HRM = human resource management. ILM = internal labor market.

32

Fredrik Andersson et al.

Table 1.3

Explained variance by HRM components

Component 1 2 3 4 5 6 7 8 9

Fraction of variance explained

Cumulative explained variance

0.255 0.172 0.135 0.108 0.091 0.077 0.061 0.056 0.044

0.255 0.428 0.562 0.671 0.761 0.838 0.900 0.956 1.000

Notes: Variance explained by relative weights of each factor’s eigenvalues from a principal. HRM = human resource management.

1.4.2 HRM Principal Components Analysis Because firms adopt discrete bundles of HRM variables, we anticipate a high degree of multicollinearity across the nine underlying HRM variables. In order to avoid overfitting our regression models, we implement a principal components regression framework. First, we construct the principal components of the underlying HRM variables using eigenvectors of the correlation matrix as coefficients.7 Each component is a linear combination of the underlying variables, and we retain the combinations that capture the most variance in the underlying data and then rotate the axes to facilitate interpretation of the components. In table 1.3, we present a summary of the variance explained by the nine HRM components. The values in the table represent a proportion of the eigenvalue from each principal component. We find that there is a lead HRM component with several secondary components of lesser importance. For the subsequent analysis, we focus on the first six components, which explain 84 percent of the variance for the set of HRM variables. Table 1.4 reports the HRM component patterns for high-education workers.8 The first component, which we label “ports of entry,” corresponds to a high level of initial earnings relative to market and a high standard deviation in initial earnings. This is the lead component and indicates 7. These principal components are then ordered by variance, and the largest components are retained and then rotated to ease interpretation. Detailed descriptions of the technique are given in, for example, Sen and Srivastava (1990, 253–55) or Draper and Smith (1981, 327– 32). While this technique has found use in some of the applied statistics literature, the technique has been shown to produce poor results in certain data sets (e.g., refer to Hadi and Ling [1998] for illustrations.) 8. The first six components from the principal components analysis were orthogonally transformed through a varimax rotation. Subject to a threshold test of .50 for significance, each human resources (HR) variable has a significant loading in exactly one component.

0.20 0.13 0.05 0.90 0.02 –0.07 0.77 –0.06 0.05

Component 2: Turnover rate 0.05 0.02 0.04 –0.01 0.93 0.01 –0.01 0.66 0.14

Component 3: Wage growth 0.94 0.02 0.07 0.00 –0.09 0.01 0.31 0.35 0.01

Component 4: Hiring rate

0.01 –0.01 0.23 0.11 0.02 0.06 –0.08 0.28 0.95

Component 5: Performance incentives

0.01 0.05 –0.02 –0.08 0.01 0.99 0.00 0.00 0.06

Component 6: Early matching

Notes: Component pattern matrix from the top six components of a principle components analysis with varimax rotation. Weights ≥ .50 are in boldface. SD = standard deviation. HRM = human resource management.

0.05 0.89 0.82 0.01 –0.07 0.03 0.22 0.23 0.16

Component 1: Ports of entry

HRM component patterns for high-education workers

Accession rate Ratio of mean initial wage to market initial wage SD of initial earnings Separation rate at 2 years tenure Within-job wage growth at 2 years tenure SD of within job wage growth at 2 years tenure Separation rate at 5 years tenure Within-job wage growth at 5 years tenure SD of within-job wage growth at 5 years tenure

Variable

Table 1.4

34

Fredrik Andersson et al.

how many ports of entry are used by the firm, as opposed to hiring at an entry level and promoting from within. A high value on this component describes firms that hire workers at many different levels of experience and skill, which increases the level and variance in initial earnings. The second component, labeled “turnover rate,” reflects a high separation rate after two and after five years of tenure. The third component, labeled “wage growth” reflects high levels of within-job wage growth after both two and five years of tenure. The fourth component, “hiring rate,” reflects the overall hiring rate in 1997. The fifth component, “performance incentives,” reflects a large variance in within-job earnings growth after five years tenure, which indicates that by this point the firm has selected certain workers for career development and advancement. The sixth component, “early matching,” reflects a large variance in within-job earnings growth after two years tenure, which indicates that new hires are already selected for specific job tracks and career development. In table 1.5, to check the correspondence between the components and the underlying variables, we present the means of each component for the HRM clusters from the previous section. The ILM systems (Clusters 1 and 3) have negative averages for ports of entry (with bureaucratic much lower than performance-based). Spot-market systems (Cluster 2 and 4) have positive means for turnover and for early matching. Overall, the component scores are consistent with our labeling of the clusters. We further summarize the components by presenting component means by R&D level. Table 1.6 demonstrates that relative to low R&D firms, high R&D firms exhibit higher values for ports of entry, turnover, wage growth, hiring rate, early matching, and a lower rate for performance incentives. These differences are consistent with the suggestion that high R&D firms are more likely to implement more market-oriented HRM systems, and low R&D firms are more likely to implement HRM systems with more long-term performance incentives. 1.4.3 Worker Productivity Regressions Next, we map the HRM variables for each firm to continuous variables corresponding to the components identified in the preceding and consider the impact of these HRM components on firm performance. Specifically, we regress productivity on the principal HRM components both with and without interaction with R&D spending. We measure firm performance as log worker productivity and control for log of physical capital (in order to capture capital intensity) and product market at the four-digit SIC (in order to capture product lifespan differences). We estimate two specifications: one specification with no R&D interactions and a second specification where R&D categories (high, low) are interacted with the HRM components. We employ principal components as regressors instead of the underlying HRM variables because of multicollinearity concerns and to

120

273

235

–0.606 –0.103 0.059 –0.061 0.014 –0.131

Cluster 3: Bureaucratic ILM

Notes: See text for definition of clusters and factors. HRM = human resource management. ILM = internal labor market.

N

0.288 0.178 0.036 –0.143 –0.016 0.052

Cluster 2: Spot market with rewards

–0.088 –0.084 –0.081 0.031 –0.070 –0.071

Cluster 1: Performance-based ILM

Component means for high education HRM clusters

Component 1: Ports of entry Component 2: Turnover rate Component 3: Wage growth Component 4: Hiring rate Component 5: Performance incentives Component 6: Early matching

Table 1.5

57

0.570 0.249 –0.069 –0.265 –0.212 0.381

Cluster 4: Spot market

56

1.206 0.216 0.062 0.056 –0.261 0.058

Residual firms

36

Fredrik Andersson et al.

Table 1.6

High-education HRM component means by firm R&D level

Component 1: Ports of entry Component 2: Turnover rate Component 3: Wage growth Component 4: Hiring rate Component 5: Performance incentives Component 6: Early matching N

Low R&D firms

High R&D firms

–0.283 –0.023 –0.100 –0.143 0.058 –0.058

0.121 0.016 0.058 0.017 –0.141 –0.004

300

441

Notes: See text for definition of components. HRM = human resource management.

address latent variable issues. As a robustness check, we also present results using a continuous measure of R&D in place of the dichotomous R&D measure. We observe that several HRM components are related to worker productivity (see table 1.7). Specifically, firms with high levels of R&D investment are likely to benefit from HRM systems with multiple ports of entry, performance incentives, and lower turnover, while firms with low R&D are likely to benefit from HRM systems without early matching. Firms with multiple ports of entry, which facilitate the hiring of workers with required skills, have higher labor productivity. This effect is more important (and significant) in the high R&D firms, which supports Hypothesis 1. Performance-based pay appears to be more important in high R&D firms than in low R&D firms. Firms with higher separation (turnover) rates appear to have lower firm performance, although this is significant only for high R&D firms. The effect of turnover rate on worker productivity appears to be significant only for high R&D firms, which does not support Hypothesis 2. Because these statistical relationships do not control for firms growing or shrinking, separation rates and hiring rates may reflect poor performing firms losing workers and high performing firms adding workers. Firms with early matching or sorting of workers appear to have lower worker productivity, although this is significant only for low R&D firms. In tables 1.8 and 1.9, we examine the robustness of our results to changes in the construction of the R&D measure. Instead of the dichotomous measure used in the previous analysis, we examine if the results are robust to use of a continuous measure of R&D intensity. In table 1.8 we reestimate the model in table 1.7 where firm performance is a function of capital-labor ratio, continuous R&D, and HRM components. In the first specification, we have no R&D-HRM interactions; in the second specification, we interact HRM practices with the continuous R&D measure. In specification 1, we find very similar results to the model estimated with dichotomous

Table 1.7

High-education HRM components on firm performance

Intercept ln(K/L) C1: Ports of entry C1 × low R&D

Col. (1)

Col. (2)

2.3187*** (0.2491) 0.3004*** (0.0306) 0.0837*** (0.0272)

2.2247*** (0.2532) 0.3022*** (0.0306)

0.0577* (0.0323) 0.1397** (0.0500)

C1 × high R&D C2: Turnover rate C2 × low R&D

–0.0564** (0.0264) –0.0132 (0.0413) –0.0829** (0.0346)

C2 × high R&D C3: Wage growth C3 × low R&D

0.0137 (0.0251) 0.0014 (0.0352) 0.0307 (0.0359)

C3 × high R&D C4: Hiring rate C4 × low R&D

0.0389 (0.0262) 0.0842 (0.0540) 0.0326 (0.0297)

C4 × high R&D C5: Performance incentives C5 × low R&D

0.0284 (0.0252) 0.0124 (0.0406) 0.0614* (0.0339)

C5 × high R&D C6: Early matching C6 × low R&D

–0.0146 (0.0246) –0.0709** (0.0355) 0.0450 (0.0351)

C6 × high R&D R2 N

0.66 760

0.66 760

Notes: Dependent variable is log worker productivity. Both specifications include controls for four-digit SIC. Standard errors in parentheses. HRM = human resource management. *Denotes significance at the 10 percent level. **Denotes significance at the 5 percent level. ***Denotes significance at the 1 percent level.

38

Fredrik Andersson et al.

Table 1.8

High-education HRM components and continuous R&D on firm performance

Intercept ln(K/L) R&D (continuous) C1 C1 × R&D C2 C2 × R&D C3 C3 × R&D C4 C4 × R&D C5 C5 × R&D C6 C6 × R&D

Col. (1)

Col. (2)

2.3631*** (0.2506) 0.2979*** (0.0306) 0.0332 (0.0219) 0.0794*** (0.0273)

2.2492*** (0.2580) 0.2966*** (0.0308) 0.0297 (0.0225) 0.1664*** (0.076) 0.0291 (0.0246) –0.0543 (0.076) –0.0004 (0.0208) 0.0541 (0.057) 0.0193 (0.0210) 0.0176 (0.044) –0.0108 (0.0199) 0.1021** (0.052) 0.0277 (0.0193) 0.0478 (0.048) 0.0236 (0.0141)

–0.0558*** (0.0264)

0.0111 (0.0252)

0.0358 (0.0262)

0.0337 (0.0254)

–0.0149 (0.0246)

R2 N

0.66 760

0.66 760

Note: See table 1.7 notes. *Denotes significance at the 10 percent level. **Denotes significance at the 5 percent level. ***Denotes significance at the 1 percent level.

R&D. However, when we examine the interactions in specification 2, we find differences between the continuous and dichotomous models. In the continuous R&D model, the turnover HRM practice (Component 2) is no longer significant, but the performance incentive practice (Component 5) is significant. The differences in the interaction terms in the continuous and dichotomous models suggest that the relationship between R&D intensity and

The Effect of HRM Practices and R&D Investment on Worker Productivity Table 1.9

39

High-education HRM components and R&D interactions on firm performance Col. (1) Intercept ln(K/L) R&D (continuous) × low R&D (continuous) × high C1 × low R&D C1 × high R&D C2 × low R&D C2 × high R&D C3 × low R&D C3 × high R&D C4 × low R&D C4 × high R&D C5 × low R&D C5 × high R&D C6 × low R&D C6 × high R&D R2 N

2.2936*** (0.2643) 0.3003*** (0.0306) 0.0357 (0.0247) 0.0448 (0.0467) 0.0552* (0.0324) 0.1247** (0.0524) –0.0125 (0.0413) –0.0823*** (0.0346) –0.0025 (0.0353) 0.0293 (0.0360) 0.0882 (0.0541) 0.0254 (0.0303) 0.0206 (0.0411) 0.0611* (0.0343) –0.0716*** (0.0356) 0.0459 (0.0351) 0.66 760

Note: See table 1.7 notes. *Denotes significance at the 10 percent level. **Denotes significance at the 5 percent level. ***Denotes significance at the 1 percent level.

HRM practices are not linear. In table 1.9, we estimate a model where we control for continuous R&D, but interact the HRM components with the dichotomous R&D indicator. We find highly similar results to the dichotomous interactions presented in table 1.7. Taken together, the two models suggest that the largest effect attributable to the interactions be-

40

Fredrik Andersson et al.

tween R&D and HRM components occurs at the high-end of the R&D scale. Overall, the regression results provide some preliminary evidence for Hypothesis 1 and mixed support for Hypothesis 2. The analysis suggests that high R&D firms benefit from HRM systems that offer multiple ports of entry, low turnover, and performance incentives, while low R&D firms benefit from HRM systems that offer multiple ports of entry and low wage variance for recent hires. The empirical results for high R&D firms are consistent with Hypothesis 1 because firms have higher productivity if they implement systems that allow hiring at many ports of entry and have tools to retain and retrain key workers. The empirical results on low R&D firms provide mixed support for Hypothesis 2. We do find that low turnover is correlated with firm performance; however, we did not hypothesize that ports of entry would play a substantial role in determining firm performance, and we do not find performance-based pay to have as large an impact as hypothesized. Although the flexibility in hiring at multiple ports of entry does not go with our strict rendition of ILM systems, this result is consistent with other empirical analyses that have found multiple ports of entry in ILMs (Baker, Gibbs, and Holmstrom 1994; Lazear and Oyer 2004). 1.5 Discussion This chapter presents evidence of the relationship between firms’ technology investment decisions, HRM practices, and productivity. We find that in the high-tech electronics industry, there is a positive correlation between performance and buying new skills (i.e., hiring at many ports of entry) for both high and low R&D firms, but the relationship is considerably stronger for high R&D firms. Interestingly, high R&D firms benefit from lower turnover and from having earnings that reward performance over a five-year period, which indicates that firms need to keep their technical workers, and the skills they have, for one project generation and into the next generation. Low R&D firms benefit from treating their workers comparably once they are hired, which indicates a positive correlation between performance and making new skills over time. In other words, high R&D firms appear to be more productive if they implement performance-based ILM systems, while low R&D firms appear to be more productive if they implement a modified bureaucratic ILM system that allows hiring at multiple ports of entry with comparable treatment once hired. A key underlying assumption of our research method is that we can infer HRM practices statistically from administrative data. Although it would be interesting to compare our inferred HRM practices to company descriptions or employee perceptions of HRM systems in place, the data do not allow this. Fieldwork observations indicate that actual practices

The Effect of HRM Practices and R&D Investment on Worker Productivity

41

and company-described HRM policies are often divergent (Pfeffer and Sutton 1999). Our inferred practices are consistent with our previous research on the semiconductor industry, where we directly observed HRM practices through field work. Using extensive data collected from in-depth site visits, Brown and Campbell (2001) demonstrate that more-advanced semiconductor fabrication plants pay higher initial wages and have shorter career paths than less-advanced plants. These results from the field work data are similar to the results in this administrative data-based project. The similarity provides evidence that inferring HRM practices from administrative data is a sound practice. A strength of this research is the richness of the data set utilized combined with interpretation through the lens of detailed industry knowledge. There is very little research that ties observed firm-level HRM systems to performance outcomes: the LEHD data allows us to analyze HRM systems and outcomes within firms for a large sample of firms, and the detailed industry knowledge allows us to understand the important issues and the context of the results. While the LEHD data provide ample sample sizes and longitudinal variation, the lack of direct measures of workers’ skills or occupation and of technological change constrains the statistical estimation and limits our interpretation of the results. Although these results must be interpreted with care, they have potential implications for understanding the mechanisms that tie together technological change and workers’ outcomes. Because technological change impacts workers at the plant level, knowledge of how HRM systems interact with technological investment to drive productivity at the plant level will inform our understanding of how labor markets work in technologically dynamic industries.

References Abowd, John, John Haltiwanger, and Julia Lane. 2004. Integrated Longitudinal Employer-Employee Data for the United States. LEHD Program Technical Paper no. TP-2004-02. Washington, DC: Longitudinal Employer-Household Dynamics Program, U.S. Census Bureau. Allen, Steven G. Technology and the wage structure. 1997. NBER Working Paper no. 5534. Cambridge, MA: National Bureau of Economic Research. Art, D., Ramanathan Gnanadesikan, and Jon Kettenring. 1982. Data-based metrics for cluster analysis. Utilitas Mathematica 21A: 75–99. Autor, David, Frank Levy, and Richard Murnane. 2002. Upstairs, downstairs: Computers and skills on two floors of a large bank. Industrial and Labor Relations Review 55 (3): 432–47. Baker, George, Michael Gibbs, and Bengt Holmstrom. 1994. The internal economics of the firm: Evidence from personnel data. Quarterly Journal of Economics 109 (4): 881–919.

42

Fredrik Andersson et al.

Barley, Stephen R., and Julian E. Orr, eds. 1997. Between craft and science: Technical work in U.S. settings. Ithaca, NY: Cornell University Press. Berman, Eli, John Bound, and Zvi Griliches. 1994. Changes in the demand for skilled labor within U.S. manufacturing industries: Evidence from the Annual Survey of Manufacturing. Quarterly Journal of Economics 109 (2): 367–97. Bound, John, and George Johnson. 1992. Changes in the structure of wages in the 1980’s: An evaluation of alternative explanations. American Economic Review 82:371–92. Bresnahan, Timothy, Erik Brynjolfsson, and Lorin Hitt. Information technology, workplace organization and the demand for skilled labor: Firm-level evidence. Quarterly Journal of Economics 117:339–76. Brown, Clair, and Benjamin Campbell. 2001. Technical change, wages, and employment in semiconductor manufacturing. Industrial and Labor Relations Review 54 (2A): 450–65. ———. 2002. The impact of technological change on work and wages. Industrial Relations 41 (1): 1–33. Brown, Clair, John Haltiwanger, and Julia Lane. 2006. Economic Turbulence: Is a Volatile Economy Good for America? Chicago: University of Chicago Press. Brown, Clair, Yoshifumi Nakata, Michael Reich, and Lloyd Ulman. 1997. Work and pay in the United States and Japan. New York: Oxford University Press. Cappelli, Peter. 1996. Technology and skill requirements: Implications for establishment wage structures. New England Economic Review (May–June):139–53. Caroli, Eve, and John Van Reenen. 2001. Skill-biased organizational change? Evidence from a panel of British and French establishments. Quarterly Journal of Economics 116 (4): 1449–92. Cohen, Wesley, and Daniel A. Levinthal. 1989. Innovation and learning: The two faces of R&D. Economic Journal 99 (397): 569–96. DiNardo, John, and Jorn-Steffen Pischke. 1997. The returns to computer use revisited: Have pencils changed the wage structure too? Quarterly Journal of Economics 112:291–303. Doeringer, Peter, and Michael Piore. 1971. Internal labor markets and manpower analysis. Lexington, MA: Heath. Draper, Norman Richard, and Harry Smith. 1981. Applied regression analysis. 2nd ed. New York: Wiley. Entorf, Horst, and Francis Kramarz. 1998. The impact of new technologies on wages and skills: Lessons from matching data on employees and on their firms. Economics of Innovation and New Technology 5 (2–4): 165–97. Hadi, Ali S., and Robert F. Ling. 1998. Some cautionary notes on the use of principal components regression. The American Statistician 52 (1): 15–19. Haltiwanger, John, Julia Lane, and James Spletzer. 2000. Wages, productivity, and the dynamic interaction of businesses and workers. NBER Working Paper no. 7994. Cambridge, MA: National Bureau of Economic Research. Handel, Michael J. 1999. Computers and the wage structure. Working Paper no. 285. Annadale-on-Hudson, NY: Levy Economics Institute. Hunter, Larry W., and John J. Lafkas. 2003. Opening the box. Information technology, work practices, and wages. Industrial and Labor Relations Review 56 (2): 224–42. Juhn, Chinhui, Kevin M. Murphy, and Brooks Pierce. 1993. Wage inequality and the rise in returns to skill. Journal of Political Economy 101:410–42. Kandel, E., and Edward Lazear. 1992. Peer pressure and partnerships. Journal of Political Economy 100 (4): 801–17. Katz, Lawrence F., and Kevin Murphy. 1992. Changes in relative wages, 1963– 1987: Supply and demand factors. Quarterly Journal of Economics 107:36–78.

The Effect of HRM Practices and R&D Investment on Worker Productivity

43

Krueger, Alan. 1993. How computers have changed the wage structure: Evidence from microdata 1984–1989. Quarterly Journal of Economics 108:33–60. Lazear, Edward P., and Paul Oyer. 2004. Internal and external labor markets: A personnel economics approach. Labour Economics 11:527–54. Levy, Frank, and Richard Murnane. 1992. U.S. earnings levels and earnings inequality: A review of recent trends and proposed explanations. Journal of Economic Literature 30:1333–81. Milgrom, Paul, and John Roberts. 1995. Complementarities and fit: Strategy, structure and organizational change in manufacturing. Journal of Accounting and Economics 19:179–208. O’Shaughnessy, Kevin, David I. Levine, and Peter Cappelli. 2001. Changes in managerial pay structure 1986–1992. Oxford Economic Papers 3:482–507. Pfeffer, Jeffrey, and Robert I. Sutton. 1999. Knowing “what” to do is not enough. California Management Review 42:83–108. Prendergast, Canice. 1996. What happens within firms? A survey of empirical evidence on compensation policies. NBER Working Paper no. 5802. Cambridge, MA: National Bureau of Economic Research. Sen, Ashish, and Muni Srivastava. 1990. Regression analysis: Theory, methods, and applications. New York: Springer-Verlag. Sturgeon, Timothy J., Frank Levy, Clair Brown, J. Bradford Jensen, and David Weil. 2006. Services Offshoring Working Group final report. Cambridge, MA: MIT Industrial Performance Center, September. Topel, Robert H., and Michael P. Ward. 1992. Job mobility and the careers of young men. Quarterly Journal of Economics 107 (2): 439–79. Ward, J. H. Jr. 1963. Hierarchical grouping to optimize and objective function. Journal of the American Statistical Association 58:236–44. Zuboff, Shoshana. 1988. In the age of the smart machine: The future of work and power. New York: Basic Books.

2 Using Behavioral Economic Field Experiments at a Firm The Context and Design of the Truckers and Turnover Project Stephen V. Burks, Jeffrey Carpenter, Lorenz Götte, Kristen Monaco, Kay Porter, and Aldo Rustichini

2.1 Introduction The Truckers and Turnover Project is a statistical case study of a single large trucking firm and its driver employees. The cooperating firm operates in the largest segment of the for-hire trucking industry in the United States, the “full truckload” (TL) segment, in which approximately 800,000 people Stephen V. Burks is an associate professor of economics and management at the University of Minnesota, Morris. Jeffrey Carpenter is an associate professor of economics at Middlebury College. Lorenz Götte is a senior economist in the Research Center for Behavioral Economics and Decisionmaking at the Federal Reserve Bank of Boston. Kristen Monaco is a professor of economics at California State University, Long Beach. Kay Porter is a business research manager at the cooperating firm. Aldo Rustichini is a professor of economics at the University of Minnesota. Earlier versions of this paper have been issued as NBER Working Paper no. 12976 (March 2007), IZA Discussion Paper no. 2789 (May 2007), and, also, under a different title— “Adding Behavioral Economics Field Experiments to the Industry Studies Toolkit: Predicting Truck Driver Job Exits in a High Turnover Setting”—as a Sloan Industry Studies Working Paper (2007). This paper was presented at the Sloan Industry Studies Annual Research Conference, held in Boston, MA, April 25 to 27, 2007, and the authors gratefully acknowledge the support of the Alfred P. Sloan Foundation for the conference. It was also presented at the Conference on the Analysis of Firms and Employees (CAFE), held September 29 to 30, 2006, in Nuremberg, Germany, and the authors gratefully acknowledge the financial support provided to the Conference by the Institute for Employment Research (IAB), the Data Access Center (FDZ-BA/IAB), the Deutsche Forschungsgemeinschaft (German Research Foundation), their research network “Flexibility in Heterogeneous Labour Markets,” the Alfred P. Sloan Foundation, and the National Science Foundation. The authors gratefully acknowledge generous financial support for the Truckers and Turnover Project research from the John D. and Catherine T. MacArthur Foundation’s Research Network on the Nature and Origin of Preferences; the Alfred P. Sloan Foundation; the Trucking Industry Program at the Georgia Institute of Technology (one of the industry studies programs initiated by the Sloan Foundation); the University of Minnesota, Morris; the Federal Reserve Bank of Boston; and from the cooperating motor carrier. We also especially thank the managers and employees of the cooperating carrier, whose involvement and active support made the project possible. The de-

45

46

Stephen V. Burks et al.

are employed, according to the 2002 Economic Census. The TL segment has a high turnover labor market for its main employee group, tractortrailer drivers, and the project is designed to address a number of academic and business questions that arise in this setting. One major part of the project matches proprietary personnel and operational data to new data collected by the researchers to create a two-year panel study of a large subset of new hires. The most distinctive innovation of this project component is the data collection process, which combines traditional survey instruments with behavioral economics experiments. The survey data include information on demographics, risk and loss aversion, time preference, planning, nonverbal IQ, and the MPQ personality profile. The data collected by behavioral economics experiments include risk and loss aversion, time preferences (discount rates), backward induction, patience, and the preference for cooperation in a social dilemma setting. Subjects will be followed over two years of their work lives. Among the major design goals are to discover the extent to which the survey and experimental measures are correlated, and whether and how much predictive power, with respect to key on-the-job outcome variables, is added by the behavioral measures. The panel study of new hires is being carried out against the backdrop of a second research component, the development of a more conventional in-depth statistical case study of the cooperating firm and its employees. This component involves constructing large historical data sets from fragmented legacy IT sources and using them to create multivariate models of turnover and productivity. Two main emphases are on the use of survival analysis to model the flow of new employees into and out of employment, and on the correct estimation of the tenure-productivity curve for new hires, accounting for the selection effects of the high turnover. The project is designed to last three and a half years, with the first halfyear for set up, and then a year for the initial intensive data collection in the panel study of new hires, in parallel with the construction of the data sets for the statistical case study and the initial generation of modeling from these data. Then there will be two years of lower-intensity work while follow-up data is collected from the participants in the panel study of new hires. The balance of the chapter is structured as follows. Section 2.2 sets the context by describing the U.S. trucking industry and the role of the TL segsigners of the field experiments used in the project thank Catherine Eckel and Kate Johnson for sharing protocol and design details from field experimental work in Mexico and for offering helpful advice on our design issues. We also thank Urs Fischbacher, the developer of the z-Tree software used in the economic experiments, for rushing development of a new version with features we needed. The views expressed herein are those of the authors, and not of the Federal Reserve System, nor the Federal Reserve Bank of Boston, nor of any of the employers of any of the authors, nor of the project sponsors.

Using Behavioral Economic Field Experiments at a Firm

47

ment within it. Section 2.3 discusses the nature of the labor market for TL drivers and why it has had a high turnover equilibrium for about twentyfive years. Section 2.4 discusses the nature of the research relationship with the cooperating firm and how it was constructed. Section 2.5 discusses the statistical analysis of historical operational and human resource data from the firm. It has two main subparts: Section 2.5.1 exhibits preliminary findings on turnover, and section 2.5.2 does the same for productivity. Section 2.6 describes the design of the panel study of new hires and has four main subparts. Section 2.6.1 describes the context of the project’s use of behavioral economic field experiments. Section 2.6.2 covers the process by which new students are trained as tractor-trailer drivers, and section 2.6.3 discusses the schedule for the data collection effort at the training school. Section 2.6.4 lists and describes the five data collection activities (three experiments and two survey-type instruments) that take place during the first two-hour session of each data collection event, while section 2.6.5 does the same for the six activities (three experiments and three survey-type instruments) during the second two-hour section of each data collection event. Section 2.7 reflects on the implications of the project for the relevant research communities and public policy. Appendix A lists the project team during the first two project years, and appendix B provides a list and time line for the main data elements being collected by the project. 2.2 The U.S. Trucking Industry 2.2.1 Segments within the Industry To a casual observer, one truck looks much like another, but in fact, the operations that provide trucking services in the United States are meaningfully differentiated from each other on several dimensions. At the broadest level, trucking operations are broken into private carriage versus for-hire carriage, based on a legal relationship: whether the carrier also owns the freight (private carriage) or is hauling it for another party (forhire carriage).1 In recent years, for-hire carriers, one of which is the focus of the present study, have typically operated about one-third of the heavy trucks in the overall U.S. fleet, but about three-fifths of the total miles run by such vehicles (Burks, Monaco, and Myers-Kuykindall 2004a).2 For-hire trucking is itself further broken into a number of distinct seg1. Private carriers are firms primarily in nontrucking lines of business who provide trucking services internally as support functions to their primary business operations. Examples might be deliveries of food by a retail grocery chain to its stores in trucks it also owns or pickups of parts for assembly at an auto plant by the auto manufacturer’s freight vehicles. 2. Heavy freight vehicles are defined here as having a gross vehicle weight (GVW) of more than 26,000 pounds, the level at which weight alone is sufficient to require the driver to hold a commercial driver’s license (CDL).

48

Stephen V. Burks et al.

ments, separated along three cross-cutting dimensions. Within each segment, interfirm competition is significant, but across segments it may be muted, or in some cases even absent. The 2002 quinquennial Economic Census, because of its use of the relatively new North American Industrial Classification System (NAICS), which is based on production process characteristics, gives a good overview of the structure of the for-hire trucking industry at this level of segmentation. For-hire truck transportation as a whole, NAICS category 484, generated $165.56 billion in revenue in 2002, or about 1.56 percent of that year’s gross domestic product (GDP).3 The first broad scale distinction within for-hire trucking is between firms that use general purpose equipment (i.e., standard enclosed van trailers) to handle general commodities and those that use specialized equipment to handle special commodities (examples of the latter would be refrigerated vans, flatbeds, tank trailers, and various other types of specialized equipment). According to the Economic Census, in 2002, general freight operations generated $111.60 billion annual revenue (67.4 percent of the total), and specialized freight had $54.01 billion annual revenue (32.6 percent of the total). A second cross-cutting broad scale distinction is between firms that make long distance intercity hauls and those that specialize in operations in and around a particular metropolitan area. In 2002, the Economic Census reports $120.21 billion in annual revenue for long-distance trucking (72.6 percent of the total) and $45.35 billion for local hauls (27.4 percent). A third cross-cutting broad scale distinction is based on the size of the typical shipment hauled, and this dimension on which firms differ is of particular relevance to the present study. It is easiest to understand this distinction by considering full-truckload service in contrast to the other two, less-than-truckload (LTL) and parcel service. At one end of the spectrum are firms like the one providing data for the current study. The archetypal TL carrier sends a driver with a tractor-trailer to a shipper’s dock to fill up the trailer with a load, typically weighing from 10,000 to 48,000 pounds.4 The driver takes the loaded trailer wherever in the United States the shipment is destined and unloads at the consignee’s dock. The driver is then dispatched empty, possibly after waiting for a while, to the next location where a full load is available for pick up. Full-truckload carriers may use specialized equipment for special commodities, but if they haul general commodities, they use general purpose equipment to maximize the chance of backhauls.5 By contrast, both parcel and LTL firms aggregate large numbers of in3. Calculation is by the authors; total GDP is from the U.S. Department of Commerce, Bureau of Economic Analysis: http://www.bea.gov/. 4. The variation is because some less-dense freight exhausts a trailer’s volume at low weight levels, while higher-density freight hits the weight limit before the volume limit. 5. That is, this is to maximize the chance of picking up a return load from near the point at which a first one is delivered.

Using Behavioral Economic Field Experiments at a Firm

49

dividual shipments collected at local terminals by local drivers into full trailer loads and move them between terminals on fixed route systems. Parcel carriers handle very small shipments (each piece typically being no larger than 150 pounds, with the average nearer to 50 pounds), and LTL carriers aggregate medium-sized shipments (widely varying, but with average size around 1,000 pounds). The Economic Census does not group parcel service firms with the for-hire trucking industry, but with air freight carriers. However, it does capture LTL and TL firms within trucking. In 2002, the TL segment dominated the general freight portion of (nonparcel) forhire trucking, with 67.9 percent of the total employment and 83.8 percent of the total revenue. If the segments of specialized freight that are primarily TL by shipment size are added to the mix,6 then TL’s share of the total employment of 1.137 million jumps to 72.8 percent, and its share of the total revenue of $124.50 billion rises to 77.1 percent. 2.2.2 Differences in the Type of Competition within Segments The differences across the segments in the operational routines needed affect the form and intensity of competition within each segment. Specifically, in the parcel and LTL segments, the need for a fixed network of freight rehandling terminals creates an entry barrier.7 While competition among parcel and LTL carriers is frequently strong, it generally takes place among incumbents. This is evidenced by the numbers of firms in the longdistance parcel and LTL segments. In parcel, there are really only four firms with full national coverage (UPS, FedEx, DHL, and the USPS).8 There are more LTL firms, but the number is still small. The 2002 Economic Census identifies eighty-nine long-distance general freight LTL firms with five or more establishments, which is the minimum number of terminals needed to give significant geographic scope; there are only fiftyseven firms with ten or more. But in TL there are essentially no entry barriers. Because TL carriers do not normally rehandle freight once it is loaded, they do not typically require terminals, nor regular route patterns, for cost-competitive operations. So a one-truck carrier can cover the entire nation, and in doing so is competitive, on a load-by-load basis, with most of the services offered by 6. Essentially, this means adding all specialized freight except household goods moving. 7. A brand new LTL carrier that wants to serve more than a single metropolitan area must create and operate a network that is of minimum size necessary to attract sufficient traffic from shippers with differing destination demands, relative to the total shipment flow densities in the geographic area it wishes to serve. But such networks exhibit strong economies of density (a combination of both scale and scope economies)—at low volumes, the average costs are high, but they fall rapidly as volume increases. The expenses of running such a network until a large enough market share is obtained to make the new network cost competitive with those of incumbent carriers are nonrecoverable (or “sunk”) if the firm exits. And the existence of a sunk cost of entry is the classic definition of an entry barrier. 8. Local parcel service is easier to enter, and there are many firms of small geographic scope.

50

Stephen V. Burks et al.

one of the TL-segment’s giants. When more complex service coordination is the key factor in market penetration, small firms can subcontract to third-party logistics providers.9 And in fact, there is a continual flow into, and out of, the TL segment, mostly by firms operating at small to medium scales. In TL, the 2002 Economic Census identified 25,831 long-distance general freight firms.10 The market concentration levels in these two segments also show the differing nature of competition. In LTL, the 2002 Economic Census puts the revenue share of the top four long-distance general freight LTL firms at 36.3 percent, while it calculates the share of the top four long-distance general freight TL firms to be only 14.7 percent. The implication of these facts is that most of TL service is what business analysts call a “commodity business” and what economists call “perfectly competitive.” As a result, the firms “at the margin,” whose choices set prices for the whole market, in TL are often not the big players, exploiting economics of scale, but may instead be the small firms in the competitive fringe of the industry segment. Their pricing is, in turn, driven significantly by the wages drivers in such firms are willing to accept. Small firms generally face more modest wage expectations from their employees than do large ones, and they also have the benefit of more personal relationships between owners, managers, and drivers. And owner-operators, who make up a significant subset of the small firms, can always choose to pay themselves less in order to get started in the business. Large firms can choose to pay a modest premium above the level set by such firms because they may have cost efficiencies in other areas, and they may be able to maintain a small price premium due to offering customers a number of different services in an integrated fashion, but if they raise their wages too high, they will make their costs uncompetitive. This industry structure sets the context for the derived demand for truck drivers in TL freight and the consequent nature of the labor market for TL drivers. 2.3 The Labor Market for TL Drivers 2.3.1 Segmented Labor Markets Emerge The American Trucking Associations’ (ATA) quarterly turnover report typically shows the average turnover rate at large TL motor carriers to be in excess of 100 percent per year (ATA Economic and Statistics Group 2005).11 Driver turnover among these carriers is an economically signifi9. Because a TL carrier can subcontract actual movements in a spot market to owneroperators, it is possible for a firm to enter TL for-hire carriage initially with zero trucks. 10. Unlike the case of LTL, because TL firms don’t have freight terminal networks, single establishment firms can be of national geographic scope, but, in fact, 997 of these had more than one establishment, which is still more firms than in the LTL segment. 11. The ATA is a federation of several separate trucking associations.

Using Behavioral Economic Field Experiments at a Firm

51

cant phenomenon—truckload carriers make up the largest segment of forhire motor carriage by employment, with approximately 600,000 drivers working at any given time (U.S. Census Bureau 2004).12 This segment of the universe of for-hire trucking firms emerged into its present form after the economic deregulation of 1980, which transformed the structure of the trucking industry. Before deregulation, the nature of the entry barriers created by government policies resulted in lots of TL output by firms using the LTL-type organization of production, with a fixed network of freight handling terminals (Belzer 1995; Burks 1999). But in the postderegulation period, carriers specialized quite strongly in one or another specific shipment size, from the smallest (parcel), through middle-sized shipments (LTL), to the largest ones (TL) (Corsi and Stowers 1991; Belzer 1995; Burks, Monaco, and Myers-Kuykindall 2004b). As the TL industry segment emerged, so did a parallel segmentation of the labor market for truck drivers (Belzer 1995; Burks 1999).13 Drivers wanting to enter employment at parcel and LTL carriers generally found job queues,14 while the labor market for TL driving jobs began exhibiting high rates of turnover. In fact, the labor market in the TL segment has essentially been in a high turnover equilibrium since soon after the end of the recessions of 1981–1982.15 2.3.2 The TL Driver’s Job To understand this situation, we start with a short description of the human capital investment needed to become a driver and then discuss the working conditions encountered by the typical driver. Driving a tractortrailer requires training for, and passing, the state-administered written and driving tests for a commercial driver’s license (CDL). Typically a high school equivalent level of literacy is required, and training begins with at least two weeks mixed between classroom work and in-truck practice. This is usually followed by a few days to as much as a few weeks of initial driving experience, which is often obtained with an experienced driver riding in the cab as a coach, while the trainee is still driving on a “learner’s permit,” before he or she has taken the final test for the CDL. While the CDL test is administered separately by each state, as of 1991 they do so under 12. The calculation is this: in the 2002 Economic Census, TL firms have 72.8 percent of the total employment of 1.137 million workers in (nonparcel) trucking, and the usual rule of thumb is that about 75 percent of the labor force employed by a TL firm is made up of drivers, the balance being made up of sales, customer service, administrative, and managerial employees. 13. In fact, the argument of the second cited work is that the labor market segmentation was itself a significant driver of the parallel industry segmentation. 14. This was especially true at unionized carriers, but was also true to some degree at nonunion ones. 15. It is an indication of the institutionalization of the high turnover secondary labor market equilibrium in TL trucking that the ATA has published its turnover report continuously since 1996.

52

Stephen V. Burks et al.

Federal standards for what must be included. It comprises both written and driving portions, and the minimum legal age at which it may be taken is twenty-one. Trucking firms generally considered a driver to be satisfactorily experienced after a year of work, so the level of human capital required places the job somewhere between unskilled and skilled, and it is best labeled as “semiskilled.” Once a driver is licensed, the key problem in retention is generally perceived to be the working conditions faced by a tractor-trailer operator in the archetypal long-haul, randomly dispatched, forty-eight-state service provided by most TL firms. In addition to the stresses of handling a big rig among swarms of cars, many drivers have very long weekly work hours on an irregular schedule. In one published survey of long-haul drivers, 21.9 percent reported working seventy plus hours each week, and two out of three drivers reported working sixty plus hour weeks (Stephenson and Fox 1996). Other surveys report similar findings (Belman and Monaco 2001). A survey of long-haul drivers in the Midwest found the median driver worked sixty-five hours, with 25 percent reporting eighty or more hours. In a twenty-four-hour period, the median hours worked was 11, median hours driving was 8.5, and median hours in nondriving work was 2 (Belman, Monaco, and Brooks 2005). These hours contrast to those in two industries in which there are occupations with similar human capital requirements, manufacturing and construction, which had average work weeks of 40.8 and 38.3 hours in 2004, respectively (Bureau of Labor Statistics 2002). A related issue is that long-haul drivers are often away from home for multiple weeks at a time, with little predictability about the date of return. In the same survey previously mentioned, only 20.7 percent of TL drivers reported that they were home almost every day, while 28.7 percent of drivers in the same study reported being home less often than once every two weeks (Stephenson and Fox 1996). In the survey of drivers from the Midwest, the median long-haul driver had last been home four days prior to the interview, though one-quarter had been away from home ten days or longer (Belman, Monaco, and Brooks 2005). A less tangible issue is that both drivers and firms like to think of CDL holders as professionals, in command of a big rig and responsible for its safe operation. But trucking is a service business, and a primary job function of the driver is to make shippers and receivers happy. The implications vary by customer shipping or receiving location, but this can place drivers somewhat lower than they might expect on the supply chain status hierarchy. Of course, not every driver in TL operations faces the same conditions. The foregoing description applies to those “running the system,” or being randomly dispatched across the forty-eight U.S. states. Some TL operations are dedicated to the service of particular large customers, and drivers in these operations have a more restricted set of pickup and delivery loca-

Using Behavioral Economic Field Experiments at a Firm

53

tions; more regular schedules, on average; and generally enjoy more time at home, as well. And some TL operations move freight between cities via trailer-on-flat-car (TOFC) or container-on-flat-car (COFC) intermodal methods. Drivers in these operations usually have regional or local runs to and from intermodal facilities and are often home nightly, or nearly so. Given these facts, a labor economist would expect to observe a “compensating differential” built into the wages of TL drivers that have the worst conditions. In other words, other things equal, TL firms should offer long-haul randomly dispatched drivers a higher earnings level than stay-athome jobs requiring similar human capital to compensate for their poorer working conditions. But dissatisfaction over wage compensation levels is frequently cited as a leading reason for TL driver turnover (Cox 2004). 2.3.3 Buying “Effective Labor” Perhaps a better way to think of the firm’s decision problem, which captures the nature of the driver labor market and the TL driver’s job, is to consider the nature of “effective labor” in this context. For a TL firm, this is the application of labor services to move trucks to and from geographically specific customer locations on the particular time schedule desired by the firm. There are three main factors that go into the cost of effective labor in this setting. One is the cost of recruiting and training new drivers to replace those who leave, to account for the lower productivity of inexperienced drivers, and also to account for any growth in business. A second is the cost of paying compensating differentials to drivers with the worst conditions to slow driver exits. The third is the operational cost of making driver working conditions better. In response to stochastic customer demands, the most efficient allocation of equipment frequently calls for irregular schedules and little time at the driver’s home terminal. When this is the case, making schedules more regular and increasing the driver’s time at home is costly. The key point is that these three cost factors can, to a significant degree, be traded off against each other, with higher expenditure in one area lowering the expenditure in another. The firm’s goal can then be construed in the standard manner: it is to find the cost-minimizing mix of these factors. Historically, the best thinking among many competing TL firms appears to be that spending more on recruiting and training is a cheaper way to get the needed units of effective labor than paying more to raise compensating wage differentials or improve schedules.16 A stable equilibrium characterized by high turnover rates defines what labor economists call a “secondary labor market” (Cain 1976; Dickens and 16. There is actually another cost factor in “effective labor” that is nonnegligible, the costs of accidents, which inexperienced drivers have at a higher rate than do experienced ones. We do not address that cost in this paper.

54

Stephen V. Burks et al.

Lang 1993).17 The persistence of the secondary labor market for drivers in TL trucking since sometime in the early 1980s has occasioned much discussion in the trucking industry trade press over the years, as well as a number of academic studies (examples include Casey 1987; Griffin, Rodriguez, and Lantz 1992; Stephenson and Fox 1996; Griffin and Kalnbach 2002; Beadle 2004). Through the ATA, the industry has commissioned significant analytic efforts to understand the management issues raised by a high turnover business model and the long-term demographic trends affecting the viability of the model (Gallup Organization 1997; ATA Economic and Statistics Group 2005). The major findings suggest that firms are aware of the trade-offs among the components of effective labor and that within this framework firms adjust to changes in the conditions of the demand for, and supply of, effective labor. It appears that as a result, the labor market as a whole also adjusts, perhaps with some lags, to such changes. A major study done by consultants at Global Insight for the ATA links the supply of truck drivers to the supply of labor for semiskilled jobs in construction because this type of work often represents the next best opportunity for likely truckers. The labor demands in these two industries are driven by significantly different macroeconomic factors. During the 1990s, when the derived demand for drivers was high, there was a modest premium— truckers’ earnings were an average of 6 to 7 percent above a position demanding similar levels of human capital in construction. The downturn of the economy in 2000 to 2001 created slack in the trucking labor market, but the arrival of low interest rates kept the derived demand in construction relatively stronger. As a result, for a few years, the average long-haul driver could expect to make less than if employed in the construction industry. By 2004, the gap had narrowed, with long haul drivers’ earnings 1.5 percent below that of construction workers (Global Insight, Inc. 2005). These facts suggest that wage levels do adjust over time to changes in the balance of labor supply and labor demand, but the persistence of the high turnover numbers shows that the levels of compensating differential being offered are not sufficient to lower turnover to the levels typical in other blue-collar jobs.18 It is well documented that the flows into and out of industry (as well as related movements of dissatisfied drivers between firms) represent a substantial cost to firms. A study by the Upper Great Plains Transportation Institute found in 1998 that replacing one dry van TL driver conservatively costs $8,234, and the industrywide cost total was estimated at nearly $2.8 billion in 1998 dollars (Rodriguez et al. 2000). The study’s authors sug17. Correspondingly, the ATA typically reports turnover rates at LTL firms to be in the 10 percent to 20 percent range, which makes them roughly equivalent in turnover to nontrucking jobs requiring similar amounts of human capital. 18. The Global Insight study used government data that does not distinguish TL from LTL among drivers for firms in long-distance trucking, but TL drivers make up the predominant share of the categories they analyze.

Using Behavioral Economic Field Experiments at a Firm

55

gested that this estimate is conservative, but it gives an idea of the magnitude of the turnover costs that TL firms must balance against the alternative costs of raising wages or adjusting operational and dispatching decisions in order to lower turnover. One might well ask whether firms have fully explored the possibilities for trade-offs among the three factors behind the cost of effective TL labor. Most firms are operating with high turnover costs and relatively lower costs for compensating differentials and operational adjustments that improve driver lifestyles. Is it possible that some large discrete shift along the frontier could move a firm out of a “local cost minimum” in this region to a different local minimum that might be lower in total costs? In fact, J.B. Hunt, then the second largest firm in the industry, engaged in a highly publicized experiment with switching from a business model with high turnover and modest wage costs to one with higher wage costs but lower turnover in 1996. It took the portion of its workforce facing the worst conditions (long and irregular dispatches) and raised wages by 35 percent, while at the same time closing down its driver training schools (Cullen 1996; Isidore 1997). The net result was a cut in both turnover and accident rates by approximately one-half (Belzer, Rodriguez, and Sedo 2002). However, the long-run net financial benefits were not as clear (Waxler 1997); most of the other large firms in the industry, including the one providing data for the present study, continue to train many of their new drivers from scratch, and nearly all TL firms use the high turnover– modest-pay-premium model. The long-run dynamics of driver labor supply and demand are made more complex by the growth of the long-haul TL industry. Between 2004 and 2014, Global Insight projects it will grow at a rate of 2.2 percent, which translates into an additional 320,000 heavy-duty long-haul new jobs. This statistic does not include the number of drivers needed to replace those who will retire during this time; the industry will need to find an estimated 219,000 additional drivers to replace the one in five drivers who are fiftyfive years old or older and are approaching retirement. Concurrently with an increase in demand for drivers, the growth rate of the overall U.S. labor force will slow from 1.4 percent to .5 percent between 2005 and 2014 (Global Insight, Inc. 2005). Another challenging trend for the industry is that to date, Hispanics, who comprise the fastest growing segment in the workforce, represent a lower percentage of drivers than they do of the overall labor supply. It is possible that the conjunction of these factors means that a secular trend toward higher prices for trucking labor has begun. This, in turn, could shift the nature of the trade-offs that firms face among the components of effective TL labor, and—along with fuel price trends and the limitations on the growth of labor productivity in trucking (Boyer and Burks 2007)—it could also dampen the long-run growth prospects of the industry (Reiskin 2006).

56

Stephen V. Burks et al.

2.4 Working with the Cooperating Firm The cooperating trucking firm is a large company of national geographic scope, with divisions that operate in several of the segments of TL trucking, including long-haul random dispatch service, dedicated carriage for large customers, and intermodal services. By revenue and employment it is among the top one hundred firms in TL. The firm began as a family-owned enterprise in the regulatory era, although it has grown through multiple acquisitions as well as internal expansions, and the original family has not been centrally involved in top management for some time. Under family control, the management culture was stable and effective, but was also, by design, relatively inward looking. It was based on longterm employment relationships with managerial and administrative ranks filled with “trucking people,” whose careers tended to be built within this single firm. A significant portion of the management started as front-line driver supervisors or, in some cases, as drivers and then worked their way up. Managers at the firm tended to learn their skills on the job and did not see much need to look elsewhere, except to service vendors who could provide expertise relevant to particular practical business problems, such as targeted marketing surveys. During the period between deregulation and the end of the twentieth century, the firm made many major and critical strategic moves, some of which were quite daring. But the decisions leading to these moves were primarily based on the vision and judgment calls of the trucking people in top managerial positions. There was little thought of broad strategic planning in the formal sense. Early in the new millennium, a new CEO, who had significant formal training in management-related areas, directed the first exercise in formal strategic planning in the firm’s history, following a process recipe provided by a major consultancy. This exercise began to increase the interest within the firm in planning as a useful activity and also increased interest in establishing the analytic foundations for planning work. The University of Minnesota, Morris, faculty began to work with the firm starting in the fall of 2002, initially on a single pilot project in the form of faculty-guided analysis by an advanced undergraduate student. The project was successful and laid the foundation for an expanding series of faculty-guided research projects over the next two years on a variety of topics. These projects operated on a gift-exchange basis: the faculty and students contributed their time as teaching and learning functions and the firm paid out-of-pocket expenses and provided access (under appropriate confidentiality restrictions) to proprietary business data. The core of the process involved selecting topics of both business and academic interest and for which advanced undergraduates could provide analyses of business use, as well as generating course-level academic output. By the third year of such projects, about twenty students supervised by six different faculty

Using Behavioral Economic Field Experiments at a Firm

57

members had done small projects on several continuing topics, from the analysis of exit interviews, to some initial turnover and productivity analyses, to work on the recruitment and retention of Hispanic employees. Within the firm, the linchpin of the process was a senior executive who had joined the firm from the outside and who had significant prior experience working fruitfully with academics. He was promoted to responsibility for a number of the aspects of human resources and driver training and moved into his new role just as the firm as a whole was opening up internally to the importance of strategic analysis. From this initial contact, UMM came to work with several other executives, at similar or higher levels of authority and responsibility, on specific projects. On the UMM side, the linchpin was an industry studies connection: the initial supervising faculty member (Burks) worked with the Sloan-funded Trucking Industry Program as a doctoral student and as a postdoctoral fellow.19 This added academic depth and polish to trucking industry institutional knowledge he had originally begun acquiring in his youth, when he worked as a tractor-trailer driver during the era of deregulation, between two bouts in graduate school. Burks’s background, along with a passion for all things trucking-related, gave him credibility with executives and allowed him to guide the UMM side of the relationship so that useful business deliverables always accompanied the academic results of interest to faculty and students. On the basis of the relationship constructed through the student projects, Burks and a second UMM researcher, biostatistician Jon Anderson, developed a small project contractually sponsored by the firm for the summer of 2004. This project began exploring the historical data retained by the firm for strategic purposes, including the analysis of the determinants of driver productivity and turnover. The larger scale design of the Truckers and Turnover Project was developed from the starting point provided by the results of this project. Burks, who devoted a sabbatical year to the project, is the principal organizer, and he has been joined in creating and developing the substantive content of the project by the coauthors of the present chapter, as well as by a number of other colleagues, who are based at several other institutions.20 2.5 Research Component One: Statistical Case Study of Historical Data Research Component One is a statistical case study of some of the historical personnel and operations data of the cooperating trucking firm. There are three interrelated parts to this component. The first is building 19. Burks was a doctoral student at the University of Massachusetts at Amherst; the Trucking Industry Program (TIP) was then located at the University of Michigan and is now hosted by the Georgia Institute of Technology. 20. A complete list of coinvestigators appears in appendix A.

58

Stephen V. Burks et al.

the data sets needed for analysis, the second is analyzing turnover, and the third is analyzing driver productivity. The goal of the first part is to take the many different data and report outputs produced by the fragmented legacy information technology (IT) resources at the firm and construct from them data sets that permit useful strategic and tactical analyses. Because the firm’s IT investments began in the early mainframe era, and those investments were focused primarily on solving succeeding generations of practical business problems, the data storage and reporting functions at the firm do not lend themselves easily to strategic use. Data set assembly, documentation, and validation are consuming, and will continue to consume, a very large part of the project’s resources. The goal of the second part is to use survival analysis to map the differences in turnover by driver group, to use hazard functions to explore the different time paths of exits by driver group, and to use Cox proportional hazard multivariate regression to analyze the interaction between the various factors that can affect exits. The goal of the third part is to use panel data multivariate regression models to map the tenure-productivity curve of new drivers as they gain experience, using a fixed effects variant to make a first-order adjustment for the impact of selection on the tenureproductivity relationship. Once the panel data model is sufficiently robust, the estimated fixed effects will then be further dissected statistically. A key (proprietary) business deliverable from this part of the project will be the assembly of the results of the turnover and productivity models to create an “expected net value of human capital” model for the investment in recruiting and training various types of drivers, who are utilized in various types of operational settings at the firm. Central academic results are expected to be generated from both the turnover and productivity models. Additionally, the analysis of Research Component Two, the panel study of new hires, will be integrated with the results of the analyses from the statistical case study. We next briefly describe the challenges and sketch a few pilot findings from the turnover and productivity analyses. 2.5.1 Initial Work on Turnover The proprietary human resource data set used for initial turnover analysis was constructed from three distinct initial data files, which share the feature that each record provides information on one driver during one calendar week. The constituent files covered different calendar periods, so we utilize the calendar window during which all three overlap, September 1, 2001, through March 31, 2005. The first file, Weekly Hires, consists of some of the data elements recorded about a driver during the week he or she is hired. Drivers who are rehired during the calendar window have more than one line in this file. The second file, Weekly Separations, contains information recorded about a driver during the week that he or she separates from the firm. Drivers who are rehired and who, as a result, also

Using Behavioral Economic Field Experiments at a Firm

59

separate more than once during our calendar window have more than one line in this file. The third file, Weekly Employment, consists of one observation in each week for each driver employed during that week. Combining all three data sets gives a complete picture, week by week, of flows in, flows out, and who is currently working for the firm. However, there are some important limitations in these data and a resulting major problem with analyzing them. The Weekly Hire and Weekly Separations data files contain a number of useful variables, including several key breakout variables, such as the driver’s division (e.g., dedicated, intermodal, system) and what kind of prior training or experience the driver had when they joined the firm.21 Unfortunately, the Weekly Employment data file is missing these key variables. This means that at the present initial stage of the analysis we don’t have this information on the drivers who do not experience either a hire or a separation event during our calendar window. And our information is incomplete for drivers who experience only a hire or only a separation event. In particular, the division to which the driver is assigned is known prospectively at the time of the hire event. But it changes later for many drivers, and we only have the updated information in the separation event record for that subset that does depart. To partially compensate for these problems, we take the following steps. Breakout variables that are of interest in the present study are carried forward to all observations on a given driver, from that driver’s hiring observation. This gives us reasonably accurate information on the previous trucking industry training or experience of each driver (because this is not information that changes with tenure). It also tells us which division of the firm’s operations a new driver is expected to be assigned to at the time of hire. Because the data on the type of work assignment is so noisy after this process, and because we would only be able to update it for those who exit, we do not pursue specific findings about the impact of the type of work on retention in the present analysis.22 A further implication of the data limitations is that we restrict ourselves in this initial work to the subset of drivers for which we observe a hiring event during our calendar window because we do not have either hire or separation observations for long-time incumbent employees and so are missing their key breakout variable values. Given an industry context in 21. Not included, on the other hand, are items such as age, gender, level of formal education, or ethnic category. 22. We experimented with the following procedure. We flowed the values from the separation observation backward, to all prior observations of that particular driver, for the variable recording division to which the driver is assigned—for those drivers who have an observed separation only. (This overwrote the forward-flowed divisional assignment data from the time of hire for those separated drivers for whom we observe the hire event.) This gives us improved information on those who separated, but at the cost that noise is differentially left in the observations on those who do not separate. The results were not credible, so we abandoned this part of the analysis until further information can be added to the data set.

60

Stephen V. Burks et al.

which there are large inflows all the time, however, this subgroup is of significant independent interest, irrespective of what might be found if a more inclusive group could be analyzed. Also, because we are not confident that we can correctly identify all the characteristics of second or later spells of employment, we here only examine the first spell of employment during our calendar window, for those drivers who have more than one observed hiring event.23 These restrictions still leave us with a lot of data: we analyze a set of more than 500,000 observations covering more than 5,000 distinct individual drivers, observed during the period from September 1, 2001, through March 31, 2005.24 Our procedure will be to first examine the survival curve for the entire set of drivers we consider here, along with the associated hazard function, which exhibits the time path of exit risk that gives rise to the survival curve. Then we will separate out the survival curves for discrete subgroups of interest and test for differences between them, and we will also examine the hazard functions for each subgroup for useful insights. It should be noted that our analysis does not distinguish between the possible different reasons for separation. In particular, of the separation events that we observe, 76.4 percent are voluntary quits, while 23.6 percent are discharges for cause, but our survival curves and hazard functions include both.25 Descriptive Results for All First-Hire-Event Employment Spells We begin by examining the survival pattern for the first observed employment spell of all drivers having a hire-event during our calendar window. Figure 2.1 displays the central results. The vertical axis indicates the percentage of the population initially entering employment that remains after each amount of time on the job, shown on the horizontal axis in weeks from the start of employment. Some key qualitative facts emerge from this picture. First, turnover rates do look extremely high. At 10.1 weeks, 25 percent of the population is gone, 50 percent have left by 29.1 weeks (the median survival time), and 75 percent have departed by 75 weeks. Second, there is a leveling off of departures in the second six months on the job, followed by an acceleration at the end of the first year. This is consistent with the fact that most of the trainees observed here who undergo the firm’s full training program sign a 23. This does not prevent us from examining rehires, as a significant number of the first spells we observe are of rehired drivers. 24. The precise number of drivers and observations is suppressed for confidentiality reasons. 25. The primary statistical methodology is survival analysis. Standard descriptive and analytical methods are problematic when the key dependent variable (here, the length of job tenure) is a time period, as ongoing spells observed at any given point in time are censored: they continue for an unknown further period. Instead, a conditional probability approach is needed to correctly take into account the statistical information contained in censored observations (Kiefer 1988; Cleves, Gould, and Gutierrez 2004)

Using Behavioral Economic Field Experiments at a Firm

61

Fig. 2.1 Kaplan-Meier survival curve: Estimates the percentage remaining from this set of drivers at each week of tenure

contract to pay back about half the cost of training (several thousand dollars) if they do not complete a year of service after training. Plus, the job options within trucking are more plentiful for drivers with a year of experience. The surprise, in fact, is that so many new drivers leave before the first year is up. Clearly, these departures cause both the firm and the drivers to incur real costs. Further insights may be obtained by examining the hazard function for this group of drivers. (See figure 2.2.) The vertical axis indicates the probability of leaving during any particular week shown on the horizontal axis, given that the driver made it to the beginning of the week.26 Here the differences in risk of departure are shown more clearly. Exit risk is highest at about six to eight weeks, which is approximately when new trainees first pull a load by themselves, without the assistance of an instructor-driver in the cab. Once drivers make it past this stage, exit risk declines sharply until the one-year mark is reached, when separation risk spikes to almost the 26. Or, to be slightly more careful, the vertical axis shows a “departure rate” because it is the conditional probability just described, divided by the number of analysis-time units contained in each unit on the horizontal axis. In our case the denominator is 1, so the rate is also a simple conditional probability. Formally, the hazard function is defined to be the ratio of the density of employment duration to the employment duration survival function, or h(x)  f(x)/S(x).

62

Stephen V. Burks et al.

Fig. 2.2 Smoothed hazard function: Estimates the rate of departure from this set of drivers at each week of tenure, conditional on survival to the beginning of the week

same level as at the beginning. Drivers who make it to the end of two years are essentially self-selected to have a high likelihood of turning out to be longer-term employees. Descriptive Results by Level of Previous Experience or Training Drivers who are hired by the cooperating firm arrive with different levels of prior training and prior experience. In figures 2.3 and 2.4 and table 2.1, the differing performance of these subgroups with respect to retention gives rise to separate survival curves and hazard functions. The best retention is exhibited by the small group (4 percent of the total) of rehires. This can be observed from the fact that their survival curve is well above the curves of the other subgroups and is quantified in table 2.1. We can see in the table that rehires have the longest time period of any group at which 75 percent still remain (almost four months), and at which 50 percent still remain (over five years). Rehires also have a retention period for 25 percent of the starting population that is so long that it cannot be meaningfully calculated in our data. This is not surprising—rehires are the self-selected subset of drivers who are not only experienced drivers, but who have worked at least once already at the cooperating firm. Having explored other opportunities, they now choose to return to this firm as their best current option.

Fig. 2.3 Kaplan-Meier survival curves by type of student: Estimates the percentage remaining from each subset at each week of tenure

Fig. 2.4 Smoothed hazard functions by type of student: Estimates the rate of departure from each subset of drivers, conditional on survival to the beginning of the week

64

Stephen V. Burks et al.

Table 2.1

Weeks of job tenure by type of student Estimated job tenure (weeks)

Drivers for whom a “hire event” is observed (N > 5,000) All drivers Rehire Experienced New students Limited experience Prior training

Percent of drivers

75% of drivers remaining

50% of drivers remaining

25% of drivers remaining

100 4 8 73 3 14

10.1 16.6 10.4 11.1 8.1 6.7

27.4 284.7 29.4 30.1 21.1 18.1

72.1 n.a.a 98.3 73.1 53.1 49.1

a

Rehires have a retention period for 25% of the starting population that is so long that it cannot be meaningfully calculated in the data.

The hazard function for these drivers is distinctive as well. It shows a modest spike in exit probability early, with falling exit risk thereafter, and also a very distinct periodicity during the first year, which likely reflects the incentive effects of the firm’s quarterly bonus system. Rehires are eligible for the firm’s quarterly bonus immediately upon starting work and also have experience with the incentive provided by the particular bonus system offered by the firm. The periodicity in the rehire hazard function suggests drivers in this group who may consider leaving during the first year are likely to wait until they have completed a quarter and have qualified for the bonus before separating. Also noteworthy, and sensible, is that there is no “first-year-effect” spike in the rehire hazard rate—this effect in the aggregate hazard function is entirely due to the behavior of other subgroups. Next consider experienced drivers. These are students who have significant levels of over-the-road tractor-trailer experience with other employers before coming to the cooperating firm. Like rehires, they only have to take a refresher training course that takes a few days, instead of the multipleweek basic training course all other drivers new to the firm are required to pass. Their retention performance is not as good as that of the rehires, but it is still well above that of the lowest groups, with 75th, 50th, and 25th percentile retention periods of 10.4, 29.4, and 98.3 weeks, respectively. Their hazard function shows the usual pattern of an early peak, with later declines, and appears to have a muted version of the periodicity seen in rehires. This would make sense, as experienced drivers are eligible for the bonus system immediately, but don’t have as much experience with its incentives as rehires. The next item to note is akin to Sherlock Holmes’s famous observation about the mysterious behavior of the dog in the night. The dog didn’t bark when it should have, and correspondingly one would expect new students with no prior background of any kind in trucking to have different (and in

Using Behavioral Economic Field Experiments at a Firm

65

particular, poorer) retention performance than experienced drivers.27 But in these data, both new students who are learning the industry from scratch and experienced drivers who are new to the cooperating firm have closely similar retention behavior for nearly the first entire year of employment. New students actually do slightly better than experienced drivers near the end of the first year. At that point, their hazard function spikes very sharply, and their performance drops below that of experienced drivers. This is likely associated with the facts that their training contracts are completed and they then have enough experience to easily switch trucking firms if they desire. Because new students are by far the largest group (73 percent) of drivers for whom we observe a hire event, their behavior is very important in determining that of the entire aggregate driver population. Thus, the size of their initial aggregate spike in exit risk, as well as that after a year of service, both strongly shape the aggregate survival curve and hazard function. As it turns out, a Chi-square statistical test of the significance of the difference in overall survival performance between new drivers and those with experience at firms other than the one providing the data shows that experienced drivers do better overall, at the 5 percent significance level ( p  .018). But, as table 2.1 shows, the effect is all driven by the one-year exits of new drivers, and the magnitude of the effect is much smaller than the difference between either of these groups and rehires.28 For instance, 50 percent of the rehire group is estimated to still be at work for the cooperating firm 5.48 years after the hire event we observe, while for drivers with experience at other firms, it is only 6.8 months, and for new students it is essentially the same, at 6.9 months.29 At longer durations of employment, we see a modest difference: 25 percent of the drivers with experience at other firms still remain at 22.6 months, while it is only 16.9 months for the same proportion of new drivers. Last, consider the retention performance of the two final groups: drivers with some prior experience and those with some prior training. Both these groups are identified by the driver recruiting staff at the cooperating firm as having some background in trucking, but not enough to qualify the student to take only the short training course for fully experienced drivers. To extend the previous allusion, here is a dog barking loudly—these two groups do quite badly, by comparison to students wholly new to trucking. The job tenure lengths for the retention of the 75th, 50th, and 25th percentiles of 27. The mysterious behavior (in “The Silver Blaze”) was that the dog did not bark when someone removed a valuable race horse from the barn, which was a clue to the thief’s identity. 28. The pairwise differences between rehires and new drivers, and between rehires and experienced drivers, are both significant—the Chi-square p-values for Type 1 error are zero to four decimal places. 29. The base time unit for the statistical analysis is weeks, so months are everywhere calculated as weeks divided by 4.33.

66

Stephen V. Burks et al.

students with limited driving experience is 1.87, 4.94, and 12.25 months, respectively. This tells us that only 25 percent make it to the completion of their one-year-service-after-training employment contract; the other 75 percent are incurring a multithousand dollar debt in order to leave early.30 Students with only some prior training, but no prior experience, do even worse, with retention periods for the 75th, 50th, and 25th percentiles of only 1.58, 4.18, and 11.33 months, respectively. So less than one-quarter of these students complete their training contracts. (The difference between these two groups is significant by the Chi-square test, at the 5 percent level [ p-value of .045.])31 Why should these students be at the bottom of the performance ranking when normally prior training or experience would be expected to improve retention? A reasonable hypothesis is that it has to do with the distinctive characteristics of a high-turnover, secondary labor market. In this type of market, there is always demand for drivers at some job or other. So someone with prior experience of any kind, as well as the graduates of any of the many commercial driver training schools, can get some job, as long as they have a CDL. It may not be a very desirable job, but it is possible to accumulate experience if one is willing to put up with some of the poorer working conditions available in an industry segment known for having poor conditions on average. In this context, coming to the cooperating firm and being willing to assume the debt contract that accompanies the full training program is a bad signal. There may be many specific reasons outside a prospective driver’s control that lead to such a decision. For example, the student could have experienced some kind of family event that stopped his or her prior training before the CDL exam or caused him or her to quit a prior job quickly. But, on average, students with some prior training or some prior experience are likely either to be job switchers who just couldn’t do better for the time being, but who will be looking to leave as soon as possible, or to be job candidates who were unsuccessful at someone else’s training course, or were otherwise judged inadequate by other firms. Either of these reasons means the student is more likely to depart. 2.5.2 Pilot Work on Productivity The pilot work on productivity utilized a different set of data files from the cooperating firm than did the turnover work described in the preceding section. We began with two data files, one containing basic information (especially hire date and separation date, if any) on all the drivers who had separated during the period of one year (for example, in some of the pilot 30. Except for those who are hired by a rival firm that is willing to pay off their indebtedness—something which is known to occur in this labor market. 31. The pairwise differences between either of these groups and any of those with better retention performance is highly significant—the Chi-square p-values for Type 1 error are zero to four decimal places.

Using Behavioral Economic Field Experiments at a Firm

67

work we used 2003), and the second, extracted at the end of that year, containing similar information on all currently employed drivers. Then two separate additional files containing demographic information, and racial and ethnic identity from voluntary Equal Employment Opportunity Commission (EEOC) employee disclosure forms, were added. Merging these using the internal employee number (driver number) as an identifier immediately caused problems. It turned out that driver numbers are not unique, but are recycled on a regular basis, so we had to delete some duplicate cases that really represented different drivers.32 “Hire date,” a key variable for survival analysis also turned out to be problematic. As one might expect in a high turnover setting, a small but significant number of drivers become reemployed, some having as many as four or five successive employment spells. The problem was that drivers gone for less time than some threshold (six months at one point, but varied over time) kept their original hire date, while those gone longer were assigned a new one. The latter fact made it impossible to distinguish rehires from new drivers with recycled driver numbers. To do a productivity analysis, the key addition to the records already described was information from the firm’s payroll records, which provide a week-by-week compilation of the items added to (or deducted from) each employee’s pay, with each such transaction constituting a line of data. The taxes and fringe benefit co-pays were in a separate data source to which we did not have access, but even so the initial files had as many as forty-four transactions per driver per pay period, with more than one million lines of data per file. We proceeded to document the different variables that contained coded information about the driver’s work assignment and pay structure, consulting subject-matter experts at the firm regularly. Each variable could take on multiple values, the meanings of which to some degree changed over time as operational needs changed. In addition, we began to document all the meanings of the values of the key variable specifying what type of transaction each line of the payroll file contained. There were several hundred distinct values of this variable, including values denoting several different types of mileage pay, dozens of types of lump sum pay for specific tasks, dozens of types of pay advances and pay deductions, and so on. After documentation, we next “rolled up” the payroll file. We sorted the file by driver and pay-week and then accumulated all the transaction-level information we were interested in having on a weekly basis into new variables so that the last transaction in each driver-pay-week record contained cumulative information for the week. The kinds of information in the re32. For the pilot work, we did not want the responsibility of making use of social security numbers, although a secure method for making use of the relevant identification information has been developed for later work.

68

Stephen V. Burks et al.

sulting records included such key items as the total (paid) miles, and the amount paid for them, and the total number of dispatches. Also included was information on various kinds of ancillary activities when they generated a pay transaction, such as paid customer stops, pay supplements for very short runs, paid maintenance delays, and so on. The payroll data thus provides a very rich set of information about what each driver does during each week. However, the payroll file records what drivers are actually paid for, which is in general a subset of what they actually do. So, for instance, the first pickup stop and first delivery stop on each loaded dispatch are not separately compensated. Extra pickup or delivery stops are paid when they occur on long-distance random dispatch loads, but only some of the time when they are on a scheduled run dedicated to a particular customer that is engineered to have multiple stops. Most drivers are primarily compensated by the mile, and these drivers are paid miles for all their dispatches, which normally includes loaded miles, plus miles pulling an empty trailer, repositioning for a new load, and also any bobtail miles (i.e., without a trailer). However, drivers generally run more miles than those for which they are paid. Paid miles are based on a least-distance routing algorithm, which is historically standard in the industry but which undercounts the actual miles by several percent (recent guesstimates by managers at our firm for the average undercount range from 4 percent to 6 percent).33 Despite these limitations, the payroll data provide a very useful starting point for the productivity analysis.34 Descriptive Productivity Results for Inexperienced Long-Haul Random Dispatch Drivers We began our pilot work with a subset of drivers for the years 2002 and 2003. The subset is those drivers who were inexperienced at hire (i.e., those who had to take the full training course offered by the firm), who were assigned to drive solo (as opposed to in a team) on long-haul random dispatch runs, and who were in their 5th week to 156th week of tenure with the firm.35 This gave us more than 100,000 pay-week observations on more than 33. This is, in part, because the standard algorithms are to and from standard reference points, and given the circuity of the road network, this undercounts actual miles on average. It is also because drivers are responsible for selecting a practical route for a large loaded tractor-trailer, which is often more circuitous than the least-distance version. In addition, drivers may choose to deviate for other reasons (for example, to run on a turnpike where the salt trucks will be out at night in the Pennsylvania mountains in winter, as opposed to a nontoll highway on which such services are more uncertain), as long as they don’t exceed certain percentage standards for excess miles and meet delivery schedules. 34. For later work, it is expected supplemental information will be added from a separate operational events data set also maintained by the firm. It is not the place to start because it has its own limitations and also because it is about an order of magnitude larger than the payroll data set. 35. Drivers begin receiving mileage pay when they first pull a load on their own, without a trainer in the truck with them, and the earliest this occurs is about the fifth week.

Using Behavioral Economic Field Experiments at a Firm

69

2,000 drivers. Examining the key dependent variable, miles per week, we observed very high variance (see figure 2.5). In particular, there were negative values and also very high positive values. The former turned out to be due to mistaken pay being charged back against a driver’s earnings and the latter to a small number of drivers from the firm’s early days who were permitted to accumulate vacation earnings over several years and were being paid upon retirement. We decided to trim the extremes and had to choose whether to leave in zero-miles weeks or use only positive-miles ones and what upper bound to use. The actual maximum number of miles that a solo driver could legally run during this period, given state speed limits and Federal Hours of Service Regulations for operators of commercial vehicles, was about 4,000 per week. But during at least part of this period, until the practice was ended, some drivers at the firm were paid for their runs only after they submitted completed paperwork for each dispatch. This meant that if they held their paperwork they could have one (or even two) weeks in a row with zero paid miles and then a week with very high miles. We decided to trim only the negative values, leaving zero-miles weeks in, and also trimmed values over 6,500 after looking at the distribution of the upper tail. Further examination showed that almost 20 percent of our observations were of zero-miles pay weeks. So we first trimmed out all the pay-week ob-

Fig. 2.5

Miles per week by week of driver tenure

70

Stephen V. Burks et al.

servations that were associated with any payroll transaction that could exogenously cause the driver to either miss work or be paid on a nonmileage basis. This included pay weeks with disability pay, vacation or holiday pay, salary (sometimes paid to driver-trainers), lump-sum training pay, and the like. Then we discovered that the payroll system was generating dummy paychecks for drivers who had separated from the firm, for several weeks after separation when the driver left owing money, for example, for things like cash pay advances or purchases at a store at a company terminal. When all of these cases were trimmed out, we reduced the number of weeks with zero miles substantially, but 6 percent of our observations remained with zero miles per week. Figure 2.6 exhibits a simple descriptive version of the tenureproductivity curve for this subset of drivers. Even after all the trimming, the remaining weeks with zero miles affect the mean values quite significantly. Without zero-miles weeks, the initial increase to full productivity is achieved at about nine months, whereas with zero-miles weeks it is nearer to a year. There is a sharp drop in the curve at one year of tenure with zero miles included. This is undoubtedly related in some way to the fact that drivers with one year of experience can more easily switch firms and also

Fig. 2.6 Average miles per week, by week of tenure, with and without zero-miles weeks

Using Behavioral Economic Field Experiments at a Firm

71

to the fact that most of the drivers in this subset, all of whom have taken the firm’s training, assume a debt of several thousand dollars for its cost, which is forgiven at the end of a year of service after training is completed.36 The balance of the pilot analysis keeps all the zero-miles weeks in the data set, but a goal of the full statistical case study is to dig deeper into this phenomenon and develop better evidence on whether they all should be included when analyzing productivity. The Impact of Selection on the Tenure-Productivity Curve A major goal of the statistical investigation of productivity is to analyze the true causal effect of increasing tenure on expected miles per week (the treatment effect), while accounting for any impact the high turnover rate might have (the selection effect). A priori, a reasonable hypothesis would be that drivers with lower productive capacity would be more likely to leave at any given level of tenure. In order to test this hypothesis, we start by running a fixed effects panel data ordinary least squares (OLS) regression model, with total miles per week as the dependent variable. We use all the independent variables from the payroll data set that plausibly measure exogenous factors that affect productivity. These include driver tenure (a linear term and as many higher-order terms as prove significant) and the number of dispatches (linear term plus those higher-order terms that prove significant). We also include variables such as the number of short-haul pay supplements, the number of paid maintenance delays, as well as dummy variables for each week of the calendar year (to capture any time-period effects or time trends), and a dummy variable for each terminal at which drivers are based (to capture any geographic effects of the home base). We use robust standard errors. The pilot version of this model has an adjusted R2 of .66. The fixed effects model constrains the coefficients on all the independent variables to be the same across all drivers, but permits each driver to have his or her own regression plane intercept, or constant. This constant, or fixed effect, which is estimated by the regression, may be thought of as a measure, specific to the model and the data, of the degree of “job match” between the driver and his employment at the firm. In the context of the model, it is the number of miles which the driver “brings to the job each week” (which can be positive or negative), according to the model estimate. Allowing this specific flexibility in the regression model provides a statistical adjustment for the relative speed with which drivers of high and low job match turn over, in predicting the miles each driver will operate per week. So we take the predicted values from the model and compute the average of these values over each week of tenure on the job. These averages, when 36. New drivers also earn their first week of paid vacation at this point, but that cannot be the reason for the drop in the averages, as those weeks have been removed from the data.

72

Stephen V. Burks et al.

Fig. 2.7 The tenure-productivity curve with (AvgMiHatFEStd) and without (AvgTotMiWkW0) the fixed effects correction for selection

graphed, produce a “selection-corrected” tenure-productivity curve. In figure 2.7, we compare this new tenure-productivity curve with the simple descriptive version of the same curve we exhibited in figure 2.6.37 When both curves are level and the selection-corrected tenureproductivity curve is below the old curve, the graph suggests that the true effect of tenure on productivity is smaller than it appears in the simple descriptive case. This implies that drivers with poor job matches are leaving differentially faster than those with good ones, which is in accord with our hypothesis. By contrast, when the selection-corrected tenure-productivity curve is either rising more rapidly than the old curve, or is above it when they are level, it says that the true effect of tenure on productivity may be larger than it appears in the simple descriptive case. This could imply that drivers with good job matches are leaving differentially faster than those with bad ones. The pilot results shown in figure 2.7 clearly show that our initial hypothesis is true from about the ninth or tenth month on. But fig37. Both curves are for the case in which zero-miles weeks that cannot be specifically explained away are retained in the data set.

Using Behavioral Economic Field Experiments at a Firm

73

ure 2.7 is ambiguous about whether this is also true early in the tenure of new drivers, when the firm has its highest rates of separations. A major goal of the full statistical case study is to clarify these pilot results. 2.6 Research Component Two: Panel Study of New Hires Research Component Two is a study of 1,069 new driver-trainees who were among those recruited by the firm to start their education at a specific training school operated by the firm. The basic design of the panel study is quite straightforward in conception, although it is quite labor intensive and costly to carry out. A large amount of data is being collected on each driver trainee, starting with an initial contact while each was in the first phase of training, and then continuing with follow-up data collections over two years of the trainee’s work life at the firm, or until the trainee exits the firm, whichever comes first. The follow-up data collections include these elements: (1) a follow-up paper survey for the driver mailed to his or her home every six months, for two years, as long as they stay employed by the firm; (2) an exit survey for the driver mailed to the driver’s home soon after their separation, if it occurs during the first two years; (3) a weekly survey with two questions sent over the satellite unit to the driver’s truck;38 (4) an initial survey (and consent form) for the driver’s spouse or significant other, asking about family/ work-life issues, and mailed to the driver’s family soon after the driver entered the study; (5) a follow-up survey to the driver’s family mailed every six months up to two years; and (6) an exit survey for the driver’s family mailed upon driver separation, if it occurs during the first two years. As subjects are informed as they enter the study, a cash gift of $5 is included in each survey mailing, with the goal of increasing the response rate. Finally, the drivers’ on-the-job performance data will be collected as part of future updating of the master data files for the turnover and productivity studies of Research Component Two. 2.6.1 The Use of Behavioral Economic Experiments A central project design goal is to perform a multivariate statistical analysis of the relationship between all the factors that are being measured and the success on the job of the trainees, where employee success is measured first by the length of time they are retained by the firm and, second, by their productivity on the job.39 While the researchers (and the firm’s managers) start with a number of hypotheses about what might matter in predicting each of these outcome variables, this research component is essentially a 38. The two questions are: “How happy are you with your job right now?” (Likert scale response), and “How many miles do you expect to run next week?” 39. The analysis has the potential to be extended to include safety performance, but that is not part of the present project.

74

Stephen V. Burks et al.

large-scale empirical investigation, designed to let the data tell us which factors matter statistically and which do not. In this regard, it is particularly exciting that we have the opportunity to employ both traditional measurement instruments and a selected set of behavioral economic experiments. There are many reasons one might have for doing behavioral economic experiments in the field. But one cross-cutting categorization is whether the investigation is primarily about treatment effects or about differences in individual characteristics. A treatment effect is exactly what it sounds like: analogous to studying the aggregate differences in the symptoms of ill patients who randomly received a specific medical treatment, as compared to those who randomly did not, an economic treatment effect is the difference in aggregate behavior across two variants of the experimental setting.40 An example is the difference in average transaction prices between traders in a pit market and those in a double-auction market for the same commodity and with the same valuations and costs. However, simple behavioral economic experiments can also be thought of as measurement tools for the characteristics, including the preferences, of individual subjects, as argued in Camerer and Fehr (2004). This is the approach that fits the specifics of the institutional setting of the Project. We anticipate that the results we accumulate will provide significant evidence on the relative utility of conventional and experimental measures of individual characteristics in predicting on-the-job outcomes and in complementing or substituting for each other. One methodological point should be noted. We plan to look at the relationships among the various measures we are collecting on each subject, as there is little evidence in the literature on this topic for many of our measures, let alone evidence using the subject population from which we are drawing our participants. If this were our only goal, it would be important to vary the order in which the different measures are implemented during the initial data collection, as it is quite possible that order effects could be important for some of these relationships. Order variations are a standard feature of many experimental economic designs when individual subjects take part in more than one experiment. However, because the central design goal of our project is to examine the predictive power of the various measures with respect to individual on-thejob outcomes, a countervailing methodological need is to present—as nearly as possible—exactly the same set of stimuli to each subject so that the relationship to the outcome variables is always the same. In addition, the complexity of administering the data collection process at the field site made it essentially impossible to vary the order, once we found a sequence 40. Of course, one of the key features which makes experimental evidence the scientific gold standard is that random assignment to treatment and nontreatment groups means that any causal factors not directly controlled for in an effective experimental design wash out of the results because they equally affect both groups.

Using Behavioral Economic Field Experiments at a Firm

75

of data collection events that fit the schedule. For these reasons, the sequence of measurements in the initial data collection was not varied. The initial data collection process took place on a Saturday in the middle of the first two weeks of the training process for new driver trainees. We next describe the training process to provide context for understanding where the initial data collection fits in the new driver’s work life. 2.6.2 The Driver Training Process The first two weeks of training for every inexperienced driver take place at one of the driver training schools maintained by the firm. The initial training includes a large dose of classroom work in which students are introduced to the firm and learn essential facts about the equipment they’ll be operating, the regulations governing commercial vehicle operators and operations, map reading and course-plotting, and the safety rules and procedures specific to the firm. It also alternates hands-on training time between truck simulators and an actual Class 8 tractor-trailer.41 Trainees first drive bobtail (tractor with no trailer), and then with an empty and a loaded trailer, on the school’s property. Once the trainee has qualified for a CDL learning permit and is judged ready by the instructors, he or she quickly begins driving on the actual roads and highways surrounding the school, with a driver-trainer in the right seat of the tractor. The simulators have two purposes. One is to speed familiarization with the basic features and operational characteristics of the Class 8 tractortrailer. An example is learning how to correctly shift a ten-speed transmission, which requires frequent double-clutching, coordinated at first by conscious attention to engine RPMs in comparison to road speed, especially for downshifts. Another basic feature drivers are first exposed to in the simulator is how to maneuver around corners on city streets with a rig that is about 65 feet (just under 20 meters) long, and that includes a trailer that is 53-feet (16.2 meters) long. A second main function of the simulators is to give drivers practice at responding to dangerous settings that could never be practiced in real life, such as how to avoid going off the road in response to a blow-out on a steering axle tire, first on dry pavement, and then on glare ice on a freeway in the middle of car traffic. Trainees who complete the initial two-week training process are officially hired by the firm on their date of completion, after passing a basic skills qualification test. Once hired, the driver goes back to his or her home base at one of the firm’s terminal locations. While the content of the CDL 41. By definition, Class 8 vehicles that can operate on the National Highway System (NHS) have a maximum GVW of between 33,000 pounds and 80,000 pounds. The units utilized by most truckload carriers, including the cooperating firm, are at the top of this range, and have a maximum GVW of 80,000 pounds. (The NHS is a large subset of all U.S. highways designated by the U.S. Department of Transportation (USDOT), on which federal size and weight standards prevail.)

76

Stephen V. Burks et al.

exam is federally regulated, the exact rules governing driver training vary from state to state.42 So depending on their proficiency level and on the state in which they will be based, trainees may also take the CDL exam at the training school, or they may do so later, at a location nearer their home base. Whether they have the CDL yet or are still using a learner’s permit, all new drivers have a second training phase, during which the trainee works for between one and three more weeks, actually hauling freight from their home base, but with a certified driver trainer in the right seat of the tractor. When the trainer judges the trainee ready (and he or she has acquired a CDL), the new driver is assigned a tractor of his or her own and goes to work on his or her own. As is mentioned in previous sections, it is typical for the TL carriers that train all or most of their new drivers to offer the training on a credit contract. The contract specifies that the trainees assume a debt of several thousand dollars, which is approximately the market value of the training. According to the contract, the firm provides both phases of the training, including access to the classroom, instructors, and trucks, plus needed classroom supplies. For the first phase of training, transportation to and from the training school, a hotel room while at training, and lunches during the training day are also covered. The trainee owes no payments on the credit extended as long as they stay employed by the firm, and the debt is fully discharged if the trainee completes a specific period of service after the point they begin working on their own (either one year, at many firms, including this one, or eighteen months at some others). Trainees who do not complete the period of service, however, become legally liable for repayment of the amount of credit extended. 2.6.3 The Initial Data Collection Process The initial data collection process was piloted in October through December, 2005, and went to full operation on twenty-two Saturdays during January through August, 2006. The school starts a class into the initial twoweek training process (described in the preceding section) every week on Sunday (thus, two classes are in residence at any given time). So trainees in target classes received a one-page flier advertising the study as part of their orientation on their first day at school. The flier explained the opportunity to take part in the University of Minnesota study the next weekend and that trainees who volunteered could expect to earn cash by participating. At the beginning of the class day on the Saturday of data collection, Burks conducted an informed consent process in which he introduced himself as a former driver turned university researcher and explained the goals and 42. The CDL comes in three categories, “Class C” through “Class A,” and for freight vehicles are differentiated primarily by the ascending maximum GVW of the trucks the bearer can operate. The Class A CDL is required to operate Class 8 vehicles on public roads.

Using Behavioral Economic Field Experiments at a Firm

77

procedures of the study to the potential subjects. The study is governed by the standards of the Human Subjects Protection Committee of the University of Minnesota’s Institutional Review Board (IRB), so a formal consent document was used which spelled out the risks and benefits of taking part in the study.43 The data collection at the training school was set up as two, two-hourlong blocks, spent doing tasks with the researchers, either on computers or with paper and pencil, with a short break in between. Training classes range from thirty-five to seventy students, and the largest group that could be accommodated at one time for data collection was thirty-two subjects, so each class was broken into two groups. The first group worked with the researchers from early to late morning, and the second group from late morning to midafternoon. On data collection days, the class day was structured so that those potential subjects who chose not to take part did not have extra training available, but instead could spend the extra time in the break room. The buses that transport students to and from the hotel bring everyone at one time in the early morning and take everyone back at one time in the afternoon. Given the monetary compensation being offered, the relatively low opportunity cost of taking part, and the credible guarantee of confidentiality from the University, 91 percent of the trainees offered the opportunity chose to join the study. Initial data was collected from 1,069 participants, with clean information on 1,036 of these.44 Except for one instrument (the Educational Testing Service’s test of Quantitative Literacy, labeled “Numeracy” in our schedule), all the instruments were administered on a wireless network provided by the cooperating firm of refurbished Dell notebook computers, with a newer Dell notebook as the master controlling computer. The software used was z-Tree (Fischbacher 2007), the toolbox for constructing computer-administered economic experiments and surveys developed at the Institute for Empirical Economics at the University of Zürich.45 The data collection took place in a temporary computerized experimental economics laboratory that the 43. This document included the following key facts: (1) the kind of data collection activities involved, both initially and later, (2) that subjects would receive $20 in initial cash “thank you gifts” for taking part, and have the opportunity to earn substantially more, (3) that the money paid to them was coming not from the firm but from two nonprofit foundations, (4) that under university rules for protecting research participants, the new data collected about them would be kept confidential from everyone except academic investigators approved by the Institutional Review Board (IRB), including specifically that it would never be available to their managers at the trucking firm, and (5) that participation was completely voluntary. 44. Due to a programming glitch, one item, Hit 15 Points, has valid information for only 893 subjects. 45. Because we are running a licensed adaptation of a nonverbal IQ instrument that requires the display of a large number of graphics files, each of which is a scanned image of a page with a pattern-matching task, we ran a beta version of z-Tree that was extended to handle graphics files. Our special thanks to developer Urs Fischbacher for providing this new version in time for our project.

78

Stephen V. Burks et al.

research team set up in one of the classrooms at the training school on the Friday before the data collection dates on Saturdays, and which was then removed at the end of the Saturday event. The team hand-built a set of cloth dividers hung from frames adapted from portable garment racks, which were installed between all the computer stations in order to separate subjects during the data collection. Because the data collection sessions exceeded the endurance of notebook PC batteries, extra power outlets were installed in the walls and a set of portable extension cords were deployed as part of the laboratory set up. The wireless network was part of a virtual local area network (VLAN) set up by the firm’s IT support personnel under a data security agreement approved by the University of Minnesota Human Subjects Committee. The VLAN separated the computers on which the confidential data was being collected from the rest of the firm’s network and provided secure data storage until the initial data collection was complete, and the project moved back to the UMM campus. Table 2.2 shows the time sequence and schedule for the actual data collection event of February 25, 2006. This class of trainees was relatively large, at fifty-nine students. Fifty-one out of fifty-six, or 91 percent, chose to take part. While it was theoretically possible for a subject to end up with only the initial $20 “thank you gifts,” we always paid at least a dollar or two more. The actual payouts depended in part on the use of stochastic devices (plastic bowls filled with colored or numbered poker chips), as described in the following account of each data collection activity. The average total earnings for the four-hour period of the initial data collection were $53, with a minimum of $21, and a maximum of $168. By design, we wanted most subjects to be pleased about their initial interaction with the researchers, given the follow-up contacts with them which are called for by the design. The latter fact—that we needed to gain not only the immediate but also the longer-term cooperation of our subjects—also affected the protocol for subject behavior during the four-and-one-half-hour sequence of data events. With undergraduate students in a university experimental economics laboratory and with a protocol that takes significantly less time than the project’s, it is not hard to ask subjects to be quiet, to not speak to their neighbors, to not do other tasks on the computers, and not to read, write, or use cell phones or PDAs. In order to provide a neutral setting, this is the behavior requested—and normally received from—typical student subjects in economic experiments, unless the treatment being investigated is one involving changing some aspect of this, such as allowing specific types of communication. We established expectations of our driver-trainee participants that were similar to this, but found that to receive willing compliance we had to allow for participants who finished ahead of others on tasks that took a long time to get up and leave the lab setting temporarily. We did this explicitly during the Numeracy instrument and during the MPQ in-

Table 2.2 Actual time 7:13 59 7:29

Data collection activity time allocations: February 25, 2006 Scheduled time

Total time

7:10 7:30

0:16

Informed consent process Start informed consent process How many people in the room? Finish informed consent process Session 1a

7:30 7:39 7:43

7:30 7:40 7:42

0:09 0:04 0:19

8:02

8:09

0:39

8:41

8:47

0:12

8:53

8:58

0:27

9:16

9:09 9:20

9:26

Check-in Information Activity 1: Prisoner’s Dilemma Computer generated: Pay on beliefs for what percentage of people will send $5 as person 1, how much person 2 sends if person 1 sends $0, and how much person 2 sends if person 1 sends $5, $1 each Activity 2: Multidimensional Personality Questionnaire No payment Activity 3: Risk/Loss Aversion 1 question, #23; green by drawing out of a bowl; everyone paid for their selection Activity 4: Demographics No payment Activity 5: Big Red Button Computer generated Release participants First person left at this time Last person left at this time Session 2a

9:37 9:45 9:47

9:40 9:45 9:47

0:08 0:02 0:09

9:56

9:57

0:47

10:43

10:30

0:25

11:08

10:55

0:08

11:16

11:05

0:12

11:28

11:28

0:12

11:35 11:40

11:40

Check-in Information Activity 1: Time Preferences 2 subjects, #12 and #19; 1 question, #18 Activity 2: Nonverbal IQ Start time of test 10:02 2 subjects, #8 and #21; pay $1 for each correct answer; pay on “pre” and “post” beliefs, $2 each Activity 3: Numeracy (Quantitative Literacy) Start time of test 10:44 2 subjects, #8 and #18; pay $2 for each correct answer; pay on “pre” and “post” beliefs, $2 each Activity 4: Ambiguity Aversion 1 question, #7; blue by drawing out of a bowl; everyone paid for their selection Activity 5: Hit 15 Points Computer generated Activity 6: Risk, Impatience, and Cooperation Survey No payment Release Participants First person left at this time Last person left at this time (continued)

(continued)

Table 2.2 Actual time

Scheduled time

Total time

Informed consent process Session 1b

12:15 12:25 12:28

12:30 12:40 12:42

0:10 0:03 0:23

12:51

1:09

0:35

1:34

1:47

0:13

1:47

1:58

0:32

2:16

2:04 2:19

2:26

Check-in Information Activity 1: Prisoner’s Dilemma Computer generated; pay on beliefs for what percentage of people will send $5 as person 1, how much person 2 sends if person 1 sends $0, and how much person 2 sends if person 1 sends $5, $1 each Activity 2: Multidimensional Personality Questionnaire No payment Activity 3: Risk/Loss Aversion 1 question, #21; blue by drawing out of a bowl; everyone paid for their selection Activity 4: Demographics No payment Activity 5: Big Red Button Computer generated Release participants First person left at this time Last person left at this time Session 2b

2:25 2:34 2:36

2:40 2:45 2:47

0:09 0:02 0:11

2:47

2:57

0:39

3:26

3:30

0:28

3:54

3:55

0:11

4:05

4:05

0:16

4:21

4:28

0:16

4:29 4:37

4:40

Check-in Information Activity 1: Time Preferences 2 subjects, #17 and #16; 1 question, #7 Activity 2: Nonverbal IQ Start time of test 2:53 2 subjects, #17 and #22; pay $1 for each correct answer; pay on “pre” and “post” beliefs, $2 each Activity 3: Numeracy (Quantitative Literacy) Start time of test 3:30 2 subjects, #17 and #16; pay $2 for each correct answer; pay on “pre” and “post” beliefs, $2 each Activity 4: Ambiguity Aversion 1 question, #22; blue by drawing out of a bowl; everyone paid for their selection Activity 5: Hit 15 Points Computer generated Activity 6: Risk, Impatience, and Cooperation Survey No payment Release participants First person left at this time Last person left at this time

Using Behavioral Economic Field Experiments at a Firm

81

strument, when differences in reading speed made the differences in completion time especially large, and we also allowed participants to read after completion if they were done early during the MPQ. We next turn to a brief description of each item in the sequence of data collection events. 2.6.4 Session One Data Collection Events The first two-hour block of data collection activities consists of three behavioral economic experiments and two more conventional measures. Prisoner’s Dilemma Our version of this experiment is a sequential and strategic one. Person 1 (the first mover) and Person 2 (the second mover) each are allocated $5. Person 1 can send either $0 or $5 to Person 2, and Person 2 can respond by sending $0, $1, $2, $3, $4, or $5 back. All funds sent are doubled by the researchers. Each subject provides their complete strategy in the game: they make both an unconditional decision for the first-mover role and a conditional one for the second-mover role (first, how to respond to being sent $0 and, second, how to respond to being sent $5, doubled to $10.) Subjects are randomly matched and their role selected by the computer after their decisions. This is a variant of the task used in Burks, Carpenter, and Götte (2006). Before each decision screen, subjects are also asked how they think other participants in the room will act in this experiment. The first question is “What percent of the participants do you think will send their $5 as Person 1?” and pays $1 if the subject is correct within plus or minus 5 percent. The second and third questions are “If Person 1 does not send/does send, what is the average that participants in this room will send back?” and pays $1 for each question if the subject is within plus or minus $0.25 of the actual average. There will be four main data items of potential interest from this experiment: first-mover choices, the estimate of first-mover behavior of others, second-mover conditional choices, and the estimate of the second-mover choices of others. We intend to analyze what other factors about the participants predict their choices in this experiment, as well as use some aspects of their choices as predictors of other experimental responses, and also of some kinds of on-the-job behavior. A potential workplace application is in predicting on-the-job cooperation. The TL driver’s job is relatively individualized, but there are two instances in which the data store from Research Component One may in the future be augmented to permit the construction of an on-the-job measure of cooperation.46 One of the benefits of the strategic form of this experiment is that the 46. See appendix B, items 4.1 and 4.2.

82

Stephen V. Burks et al.

Fig. 2.8

Second-mover strategies in the sequential prisoner’s dilemma

second-mover choices provide a fairly clear typing of the respondents, in terms of who is an egoist (never sends anything to the first mover), who is a conditional cooperator (sends back exactly what the first mover sent), and who is an altruist (always sends everything), with some gradations possible for intermediate dollar responses. As figure 2.8 shows, participant behavior is varied in this experiment, but “pure” versions of these three strategies predominate among second movers (shown by the large circles, which are sized proportionally to the number of respondent choices they capture). The three pure strategies make up 61 percent of the total responses, and most of the rest are intermediate responses. Those responses below and to the right of the 45-degree line fall in the “wingnut” category; these participants may have misunderstood the instructions, a hypothesis we can test by correlating these responses with some of our measures of cognitive performance. Multidimensional Personality Questionnaire (MPQ) The Multidimensional Personality Questionnaire (Patrick, Curtin, and Tellegan 2002), otherwise known as the MPQ, is the second measurement instrument. This is a standard personality profile that consists of eleven different scales that represent primary trait dimensions: well-being, social potency, achievement, social closeness, stress reaction, alienation, aggression, control, harm avoidance, traditionalism, and absorption. The short

Using Behavioral Economic Field Experiments at a Firm

83

version used in the study has 154 multiple-choice questions. Almost all of the 154 questions have the same four possible answers: “Always True,” “Mostly True,” “Mostly False,” and “Always False.” There is no separate payment for this survey. A principal use of this scale will be to generate a set of control variables in regression modeling of on-the-job outcomes. We also expect to look at how the traits defined by the MPQ relate to answers to the other survey instruments and to behavior in the experiments. As figure 2.9 shows, the trainee drivers are similar overall to the benchmark population presented in Patrick, Curtin, and Tellegan (2002), which is promising given that our primary goal is to use the variation across individuals in the components of the MPQ. Some initial hypotheses, for instance, are that high “achievement” orientation should be correlated with job success as a TL driver, especially given the incentive-pay method faced by these drivers. One noteworthy point is the apparent difference from the population in the factor “unlikely virtues.” The questions in this factor are set up so that only someone who is “too good to be true” will have a high score, and the authors of the instrument intend it as a check on the truthfulness of the subject’s re-

Fig. 2.9 Multidimensional Personality Questionnaire profile of subjects compared to MPQ reference population profile

84

Stephen V. Burks et al.

sponses. This MPQ factor offers the prospect that we may be able to control for this effect to some degree in multivariate analysis by using the unlikely virtues score as a regressor. Risk/Loss Aversion In the risk/loss aversion experiment, there are twenty-four questions that are divided into four blocks of six questions each (see table 2.3). There are two possible choices for each question, an amount of money received with certainty, and a 50/50 gamble that pays a higher dollar amount if “your color” is chosen and a lower dollar amount if the “other color” is chosen. To avoid any hint of experimenter control over the outcome, subjects choose which color (blue or green) is theirs for the random outcomes as their last response. Each of the four blocks of six questions follows the same format. In each block, the amount for “your color” and the “other color” do not change, and the amount of the “for sure” option increases by $0.50 per question. The design is intended to identify where the subject crosses over from the “for sure” option to the “gamble” option, relative to where a risk-neutral person would do so. After all subjects have made their decisions, one poker Table 2.3 Question no. 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24

Risk/loss aversion For sure

Your color

Other color

Win $2.00 Win $3.00 Win $4.00 Win $5.00 Win $6.00 Win $7.00 Win $0.00 Win $0.50 Win $1.00 Win $1.50 Win $2.00 Win $2.50 Lose $2.50 Lose $2.00 Lose $1.50 Lose $1.00 Lose $0.50 Lose $0.00 Win $1.00 Win $1.50 Win $2.00 Win $2.50 Win $3.00 Win $3.50

Win $10.00 Win $10.00 Win $10.00 Win $10.00 Win $10.00 Win $10.00 Win $5.00 Win $5.00 Win $5.00 Win $5.00 Win $5.00 Win $5.00 Win $1.00 Win $1.00 Win $1.00 Win $1.00 Win $1.00 Win $1.00 Win $5.00 Win $5.00 Win $5.00 Win $5.00 Win $5.00 Win $5.00

Win $2.00 Win $2.00 Win $2.00 Win $2.00 Win $2.00 Win $2.00 Lose $1.00 Lose $1.00 Lose $1.00 Lose $1.00 Lose $1.00 Lose $1.00 Lose $5.00 Lose $5.00 Lose $5.00 Lose $5.00 Lose $5.00 Lose $5.00 Win $1.00 Win $1.00 Win $1.00 Win $1.00 Win $1.00 Win $1.00

Using Behavioral Economic Field Experiments at a Firm

85

chip is drawn by a participant that identifies which question (one through twenty-four) everyone will be paid on, and then a second participant draws a colored chip to determine whether the winning color, for subjects who chose the gamble on the selected question, is green or blue. The choice blocks overlap the zero-dollar point in order to allow inference about the likelihood that the implied utility function has different slopes in the positive and negative domains. But there is no set of choices only in the negative domain because the project is relying on the long-term responsiveness of participants, and pilot work showed that a choice block fully in the negative domain was perceived as too unfriendly a task. Figure 2.10 shows the distribution of the total number of risky choices for our subject pool and shows that there is significant variation across participants, which is desirable for the purpose of using the variations across individuals to predict other behavior. Risk aversion and/or loss aversion are potentially key explanatory factors in TL firm turnover because nearly all of each driver’s pay is based on piece rates (cents per mile for a fixed number of miles for each dispatch). Paychecks can vary quite substantially from one week to the next, and, thus, there is a certain level of short-run financial risk that goes with the job. The question for the study is to what extent this form of experimental measurement of financial risk is predictive of the impact of pay fluctuations on driver turnover, as compared to things such as survey responses and credit score.

Fig. 2.10

The total number of risky choices

86

Stephen V. Burks et al.

Demographics This instrument is a compilation of thirty-seven demographic questions. The types of questions vary, but the areas that are covered include education level, languages, marital status, previous job experience, family lifestyle, racial or ethnic identity, country of birth, how many people are in the household, household income, and networking done while at training. These questions are mostly multiple choice, with a few numerical answer questions as well. There is no separate payment for completing this demographic questionnaire. Figure 2.11 shows the distribution of ages of the trainee drivers, as compared to the distribution among blue-collar workers, as computed from the March 2005 Current Population Survey. The minimum age to acquire the CDL is twenty-one, so there are no new trainees that are less than this minimum age. However, in general the trainee pool is similar in distribution to the entire population of blue-collar workers in the United States, albeit with a modest shift toward the age range of twenty-one to thirty. Figure 2.12 shows the distribution of the highest education level completed among the trainee drivers, with the distribution among blue-collar

Fig. 2.11

Distribution of participant age with benchmark

Using Behavioral Economic Field Experiments at a Firm

Fig. 2.12

87

Participant education levels with benchmark

workers in the United States as a comparison. The minimum literacy requirements for the TL driver job are undoubtedly the reason there are very few trainees with very low educational attainment, compared to the bluecollar workforce in the United States. It is interesting that the trainee group has a much higher proportion of individuals with some college, but possibly a lower proportion with a junior college or technical degree, as compared to the population of blue-collar workers. Given that someone who has acquired some level of useful education in college is overqualified for this job, this pattern is consistent with the view that completing a two-year degree is indeed correlated with job market success as fewer of these individuals are in our subject pool. A speculation about the apparently slightly higher proportion of four-year-degree holders in the trainee group, as compared to the benchmark, is that these are folks with other options who are attracted (at least initially) by the TL-driver lifestyle. The Red Button This experiment is designed to measure short-term impatience. It is the last data collection event in the first of the two, two-hour sessions, and it is followed by ten to twenty minutes of break time, during which participants can use the restroom and stretch their legs and get something from the

88

Stephen V. Burks et al.

vending machines if they wish. It is directly preceded by the demographic questionnaire, to which it is linked. This means that, unlike most of the other instruments, subjects do not have to wait until everyone else completes the prior activity, but instead each participant can start the red button task as soon as he or she has individually completed the demographic questionnaire. This setup is designed to minimize peer effects in the red button task as each subject will start the task at a different time. The red button activity takes exactly ten minutes, and the task is to wait quietly until a timer on the task screen counts down from 600 seconds to zero. As described in section 2.6.3, during each session subjects are requested in general to either work on the task at hand or wait quietly. We ask that they do not to talk to their neighbors, nor engage in any other activity, such as pulling out something to read. During the red button task, these rules are quite strictly enforced. Participants are paid $5 if they complete the task by waiting the full ten minutes. However, subjects may choose to end this experiment earlier if they wish. Each click of a button below the countdown clock on the computer screen, which is labeled “Reduce Waiting Time,” will cut the individual’s waiting time—but each click also costs them $1. The first click reduces the waiting time by five minutes; the second click reduces the waiting time by three more minutes, and the third click reduces the waiting time by an additional two minutes. Once the timer says zero, no matter how long the subject has waited, then he or she is free to get up and leave the room for break time. As figure 2.13 shows, 29 percent of the participants clicked at least once. This is actually lower than we aimed for; our initial calibration runs may have led us to set the opportunity cost of the first click too high. Patience is relevant to quit decisions, and being impatient can lead to poor job performance when a significant part of the job involves being able to wait on customer docks and then hurry up when customers are ready. Being patient in traffic is also a relevant job skill. We expect to investigate the extent to which red button choices will predict impatience on the job, as measured by quit decisions, controlling for prior earnings and prior waiting time during the earlier data collection activities. 2.6.5 Session Two Data Collection Events The second two-hour block of data collection activities contains three behavioral economic experiments and three more conventional measures. Time Preferences In this experiment, there are twenty-eight questions that are divided into four blocks of seven questions each. There are two possible choices for each question: a smaller amount of money paid sooner, and a larger amount of money paid later. Each of the four blocks of seven questions follows the same format. The amount for the higher payoff at a later date is al-

Using Behavioral Economic Field Experiments at a Firm

Fig. 2.13

89

Clicks to reduce waiting time at a cost

ways $80, and the amount for the lower payoff at an earlier time begins at $75 and decreases by $5 intervals to $45. The point at which a subject switches from the later payoff to the earlier one (if a switch is made) provides an implicit point estimate of the subject’s discount rate over that time horizon. The time frames are (1) today (Saturday) versus Monday; (2) today versus next Thursday; (3) Monday versus Monday plus one week, and (4) Monday versus Monday plus four weeks. The two matched pairs of time frames are designed to allow subjects that have different discount rates for choices with and without a front-end delay, that is, with and without an initial delay for both payments.47 After all subjects have made their decisions, a participant draws one poker chip to select which of the twenty-eight questions will be activated, and then two poker chips are chosen the same way which identify two of the subjects in the session who will be paid for their choices on that question. Payments are, of course, made on the dates requested, either in person, or with an official bank check mailed in a University of Minnesota envelope. As figure 2.14 shows, the three modes of the distribution of future 47. The contrast between choices in which one option is an immediate payment versus those in which both payments involve a delay will permit us to observe “impatience for immediate results,” often described in the literature as quasi-hyperbolic discounting, if it should appear.

90

Stephen V. Burks et al.

Fig. 2.14

Total future choices

choices are at zero, fourteen, and all, with a median of seventeen future choices out of a possible twenty-eight. There is a wide dispersion in individual responses. We expect to examine the relationship between this measure and such other measures as nonverbal IQ, risk/loss aversion, and impatience. And we will investigate the extent to which it adds predictive power to our statistical models of quits. Nonverbal IQ The IQ instrument used is a computerized adaptation of the Standard Progressive Matrices by J.C. Raven (Raven, Raven, and Court 2000); the authors created the adaptation under license from The Psychological Corporation.48 Each question is presented as a graphic image. On top is a large rectangular box containing some kind of a pattern with a piece missing out of the lower right hand corner. On the bottom are six (or eight) possible pieces that could be used to complete the image on top. Each section starts with easy images and gets progressively more difficult. The original instrument consists of five sections, each containing twelve questions. Our version is not fully comparable to the standard results for the instrument because we are actively administering only the last four sec48. Pilot work was done with two other instruments, but one had irresolvable licensing issues, and the other proved unsuitable for our setting. The first several administrations of the Raven’s instrument were done with paper and pencil, while we arranged the license for the computerized adaptation (used by permission of Harcourt Assessment).

Using Behavioral Economic Field Experiments at a Firm

91

tions, due to time constraints. The original is primarily benchmarked in an untimed format, and we cut our subjects off after thirty-one minutes, having given a warning three minutes earlier.49 The simplest way to match our data with the untimed reference benchmarks is to note that everyone in our subject pool finished the fourth section before the announcement of a time limit. Our correction is then simply to multiply our subjects’ scores on sections two, three, and four by 5/3 to scale them up to the same range as the score for untimed subjects doing all five sections.50 After both verbal and written instructions and two practice questions, subjects fill out a “confidence question” that asks them how they think they will do as compared to other subjects in the room, by placing themselves in the correct quintile of the distribution of scores. When the Raven’s task has been completed, the same confidence question is asked again. Subjects are paid an additional $2 for placing themselves in the correct quintile. In addition, two subjects are randomly chosen to be paid $1 per correct answer, for total possible earnings of $48 each for their question answers. Starting at about subject 200, we also began asking whether subjects want to find out their own score and the group average when they receive their payout.51 Figure 2.15 shows approximately how our panel compares to a standard benchmark population (citizens of Iowa City, IA, in the late 1990s.) On a scale of 0 to 60, our subject pool has a higher density between 36 and 44, and a lower one from 54 on up. The median for our subjects is about two points lower than that for the benchmark population. The distribution looks reasonable compared to groups of workers in other countries against which the regular version of the instrument has been benchmarked. The primary use of this measure in the project analysis will be as a control variable in predicting on-the-job outcomes. But we also expect to look at the relationships between this indicator of nonverbal IQ and the other participant characteristics that we measure. We will also examine how IQ, confidence about one’s performance both before and after the task, and the desire for full information about one’s performance are related. Numeracy (Quantitative Literacy) This instrument is part of the test of adult quantitative literacy from the Educational Testing Service. The full instrument consists of two sections, of which only the first section was used here, due to time constraints. The section is made up of twelve questions and subjects are given exactly twenty 49. There is one published benchmark from a French subject pool using a thirty-minute time limit, but including all five sections. 50. This approach produces unbiased estimates at the cost of greater variance. 51. The two selected to be paid for correct answers will learn their score from their payoff, so they will just get the average as new information, but these two subjects are selected after this question is asked.

92

Stephen V. Burks et al.

Fig. 2.15 Density of (adapted) nonverbal IQ instrument scores for drivers compared to density of benchmark population scores

minutes to complete the test. The test requires subjects to be able to add, subtract, compare numbers, compute a percentage, fill out a form, and to be able to read and understand a short problem, among other things. As with the nonverbal IQ, after instructions and a brief practice question, subjects fill out a “confidence question” that asks them how they think they will do as compared to other subjects in the room, by quintiles. When the numeracy task has been completed, the same confidence question is asked again. Subjects are paid an additional $2 for placing themselves in the correct quintile. Two subjects are randomly chosen to be paid $2 per correct answer, for total possible earnings of $24 each for their question answers. In addition, at about subject 200 we began asking whether subjects want to find out their own score and the group average when they receive their payment.52 The distribution is not directly comparable to the distribution for the full instrument (i.e., when both sections are administered), but we appear to be getting good dispersion in performance (see figure 2.16). The mean score is 71 percent. The primary use of this in the project analysis will be as a control variable in predicting on-the-job outcomes. The job of TL driver re52. As with the nonverbal IQ indicator, the two selected to be paid for correct answers will learn their score from their payoff, so they will just get the average as new information, but these two subjects are selected after this question is asked.

Using Behavioral Economic Field Experiments at a Firm

Fig. 2.16

93

Distribution of numeracy scores

quires continual application of numeracy skills, for example, in map reading and route planning, or in calculating hours remaining and hours coming available at specific future times, under the hours of service regulations for commercial motor vehicle operators. But we also expect to look at the relationships between this indicator of numeracy and the other participant characteristics that we measure. We will also examine how numeracy, confidence about one’s performance both before and after the task, and the desire for full information about one’s performance are related. Ambiguity Aversion Our ambiguity aversion experiment is identical to the risk/loss aversion experiment described in section 2.6.4, except for one detail: it is no longer known to be a 50/50 gamble if the subject chooses the uncertain option. Two blue poker chips and two green poker chips are placed in the bowl from which the winning color is drawn, and then out of sight of the subjects, six more chips are added that can be all green, all blue, or any mixture thereof. As a result, subjects only know that there is at least a 20 percent chance that green will be drawn and at least a 20 percent chance that blue will be drawn. All other features of the experiment are unchanged: there are four panels of six choices each, with the dollar values for certain and lottery outcomes as before. All subjects choose the certain payoff or

94

Stephen V. Burks et al.

the lottery in each question in each panel and then which color is theirs for the lottery outcomes. Figure 2.17 shows results for the first panel, when the choice is between a fixed amount ranging from $2.00 to $7.00 or a lottery with outcomes of $2 or $10. The horizontal axis shows the total number of lotteries chosen in the initial risk/loss aversion task, while the vertical axis shows the same total for the ambiguous version. Participants who made the same choices in both cases would be on the 45-degree line. The size of the circles shows the number of participants at each node of the grid, and it is apparent that many subjects did change the number of lotteries they selected in widely varying ways. We will be looking to see which subjects choose more or fewer risky options and whether their shift, if any, is related to other experimental or survey measures and to on-the-job success. Hit 15 Points This is a backward induction, or planning, task in the form of a small game between subject and computer. The computer and the subject take turns adding points to the “points basket,” and during each turn, the subject or the computer must add either one, two, or three points to the points basket. The goal is to be the player to add exactly the fifteenth point. The number of points in the points basket at the beginning of the round varies, and the computer and participant take turns going first. The first round is

Fig. 2.17

Ambiguity and lottery choice

Using Behavioral Economic Field Experiments at a Firm

Fig. 2.18

95

Success in the Hit 15 Points game

set so as to give the subjects an example of how the first stage of backward induction works. Before each decision is made by the subject as to what number of points they want to add to the points basket, the subject is asked whether he or she will win or if the computer will win. The subjects are paid $1 for each round that they win. Figure 2.18 shows the distribution of scores on this task. The median falls between two and three correct rounds of the game.53 Together with numeracy, we anticipate that this measure will be predictive of on-the-job success, and especially of high versus low productivity. The TL truck drivers have to do numerical backward induction every day, to calculate back from routing and delivery goals that are from a few hours to a few days ahead, in order to decide on the optimal course of action in the present. One regularly used application is in figuring out how many hours they need to get to their destination and how many hours they have available to drive under the regulations governing the hours of service for commercial vehicle operators. Risk, Impatience, and Cooperation Survey The last instrument used during the Saturday initial data collection events with the panel study participants is a list of attitude questions about 53. As mentioned earlier, the N is 893 for this task, as we discovered a programming error in the z-Tree code for the game after 176 subjects had already taken part.

96

Stephen V. Burks et al.

risk preferences, patience and impatience, and cooperation. It was assembled from a selection of papers in the literature that propose various surveyresponse measures for these characteristics of subjects. We intend these to also be tried out as control variables in order to see if they are useful and, if so, whether they are substitutes or complements to the experimental measures in predicting behavior on other measures and on-the-job success. 2.6.6 Other Measures of Interest The applicant information collected by the cooperating firm on trainees provides a few other data items of interest. There is an indicator of whether the participant is a smoker (whether the trainee requested a hotel room in which smoking is permitted during initial training). This may turn out to be related to risk attitudes. In addition, because the trainees apply for credit for the training contract, the firm is able to supply us with a credit score. We want to examine the relationships among our measures of risk/ loss aversion, ambiguity aversion, time preference, and impatience, on the one hand, and the credit score, on the other. We also are interested in determining whether the credit score useful in predicting on-the-job success, and, if so, whether it is a complement or substitute for our other measures. 2.7 Some Implications of the Truckers and Turnover Project The present chapter has presented the context and design of the Truckers and Turnover Project at a stage when much of the initial data collection has taken place, but while follow-up data is still being collected and before results have become available. The proof of this pudding will, of course, be in the eating, when data collection is complete and results are available. But there are three points that can be made about the project at this stage. The first two are about research methodology, and the third is about the policy significance of the subjects of the project. 2.7.1 For the Industry Studies Community First, for the industry studies research community, the project stands as a kind of “proof of concept” for adding behavioral economic experiments to the already rich mix of data collection techniques it utilizes. The Sloan Foundation’s Web site states that the Foundation’s goal in supporting industry studies is “to build a strong community of industry studies scholars who contribute to academia, to industries, and to government.” Along with an interdisciplinary approach, the defining feature of this work according to the Foundation is that “industry studies scholars develop a deep understanding of particular industries by grounding their research in direct observation” (2007). The conventional tools for direct observation range from in-depth participant observation, to shorter structured site visits and interviews, to the collection of proprietary internal data for statistical analysis. To this list, the Truckers and Turnover Project prospectively

Using Behavioral Economic Field Experiments at a Firm

97

adds behavioral economic field experiments. The key, in addition to having an existing cooperative research relationship between academics and firm(s), is to design the behavioral experiments to fit the specific institutional and business context. The new data-collection techniques must both complement more traditional approaches and also must credibly add value, from both an academic and a business standpoint. In this regard, the essentially individualized production process of TL motor freight is especially suited for an initial exploration of the uses of experiments that measure individual characteristics of employees. But one can easily imagine more complex social dilemma experiments being applied in teamwork employment settings, for instance.54 2.7.2 For the Experimental and Behavioral Economics Communities Second, for the behavioral and experimental economics communities, the project also stands as a kind of “proof of concept” for the specific manner of applying behavioral experiments to the workplace used. As mentioned in section 2.6.1, there are a wide variety of scientific motives for conducting specific types of economic experiments, but one large crosscutting categorization is whether the primary interest is in treatment effects or in measuring the characteristics of individual subjects. The latter is the purpose here, in keeping with the goal mentioned in the preceding paragraph about targeting the design to the specific institutional context. The laboratory and the field are both relevant sources of empirical data to be used in the generation and evaluation of economic theory. But it is often difficult to translate the import of findings from studies in the fully controlled, abstract, and simplified setting of the lab directly to the uncontrolled and complex world of real employment relations and production processes. The approach here is close to what Harrison and List (2004) call a “framed field experiment,” in which the experiments are similar to those that might be done in a lab with student subjects, but are instead conducted in the field setting of the driver training school, with new driver-trainees as subjects so that the framing for the (relatively abstract) experimental tasks is provided by these contextual factors. We anticipate finding some new empirical regularities when we examine the ability of our experiments to predict on-the-job outcomes, and, if so, we will have provided significant evidence for their “external validity,” that is, for the ability of the lab measurements we use to capture individual characteristics that are important in practice in a specific real workplace setting. 2.7.3 The Larger Significance of Our Subject Pool The last point we want to make is about the economic and human importance of the subjects of our study, driver-trainees in the truckload seg54. For some promising work along these lines, see the series of papers on fruit pickers in England by Bandiera, Barankay, and Rasul (2005, 2006).

98

Stephen V. Burks et al.

ment of the motor freight industry. Much has been made in the last decade of the growth of the knowledge economy and of the role of workers who have the specific education and skills to prosper in it. But what about those who do not fit this description? In 2005, fully 40 percent of the U.S. workforce had a high school degree or less, and only 29 percent had a four-year college degree or more (Mishel, Bernstein, and Allegretto 2006, 153). As the United States continues its transition from a manufacturing economy to one in which service occupations and service industries dominate total employment, nonknowledge service workers will become an increasingly important part of the overall picture. The job of TL tractor-trailer driver is an archetypal example of nonknowledge work that is important in the knowledge economy. The United States is a geographically large economy, so transportation and distribution have always been a key complement to U.S.-based production. And as the international trade in goods looms larger and larger in U.S. production and consumption, the more essential become the tasks associated with physically transporting traded goods and their ingredients here in the United States. Unlike some other types of services, the tasks involved in physically moving freight here are not subject to direct competition from abroad. According to the 2002 quinquennial Economic Census, trucking is the largest single industry in transportation services by employment, and the truckload segment of trucking had a headcount of approximately 800,000. This means that about 600,000 persons are TL tractor-trailer drivers at any one time.55 Yet the ATA’s survey results say that the typical turnover rate at TL firms is at least 50 percent per year.56 If these figures are taken at face value, it means that several hundred thousand people train for and try out this job each year, only to leave it within a few months, probably having incurred a significant debt for training that most have little hope of repaying. Our pilot findings on turnover at the cooperating firm do not contradict this interpretation of the industry facts. In this context a primary goal of the Truckers and Turnover Project is to use the best available means to measure the individual characteristics that make for success in this job. We will then ask whether the findings have actionable implications for business strategy and for public policy with respect to the labor market and education. Clearly, improving the match between job and employee by even a small amount would be a real improvement in economic circumstances for both firms and employees. It may even be that our findings will help in understanding how and to what extent the nature of these jobs can be improved, by quantifying more clearly the costs and benefits of lowering turnover. 55. This is by the rule of thumb that drivers make up about 75 percent of employment at a typical TL firm 56. The rate for large carriers typically averages over 100 percent, but most firms are small, and the rates for this group are usually between 50 percent and 100 percent per year.

Using Behavioral Economic Field Experiments at a Firm

99

Appendix A The Truckers and Turnover Project Team Project Organizer: 1. Stephen Burks, Division of Social Science, University of Minnesota, Morris; IZA; and Trucking Industry Program (Georgia Institute of Technology) Project Coinvestigators (each colleague is collaborating on at least one component of the project, some on multiple components): 1. Jon Anderson, Division of Science and Math, University of Minnesota, Morris 2. Jeffrey Carpenter, Department of Economics, Middlebury College; IZA; and Norms and Preferences Research Network 3. Andrew Clark, Centre National de la Recherche Scientifique, PSE (France), and IZA 4. Lorenz Götte, Research Center for Behavioral Economics and Decisionmaking, Federal Reserve Bank of Boston, and IZA 5. Aldo Rustichini, Department of Economics, University of Minnesota, Twin Cities 6. Kristen Monaco, Department of Economics, California State University at Long Beach; and Trucking Industry Program (Georgia Institute of Technology) On-site Research Team Members 2005 to 2006: 1. Kay Porter, Business Research Manager, Cooperating Firm 2. Adam Durand, Research Intern (fall 2005), University of Minnesota, Morris 3. William Leuthner, Research Intern (spring 2006), University of Minnesota, Morris 4. Erin Christenson, Research Intern (summer 2006), University of Minnesota, Morris

100

Stephen V. Burks et al.

Appendix B Summary List and Time Line for Data Items Collected 1.

Project Component One: Statistical Case Study

Data items here come in separate data report files, each of which has different data definitions and data errors. These are joined into a master data file by the project team, in a labor-intensive data validation, cleaning, documentation, and merging process. Initial data set covers January 1, 2002, through December 31, 2005, by construction from the individual years. Updates are planned on semiannual or quarterly basis. 1.1. Payroll-based data. Rolled-up totals of activities for which drivers are paid, by driver and week. Some pay deductions, but not all that would be of economic interest, are included. Typical items of interest include paid miles, the number of units of mileage pay (approximately the number of dispatches), and the number of times supplementary pay was received for particular work event (examples: hand unloading, weighing the rig on a scale). Includes operations-related driver characteristics (e.g., the terminal at which the driver is based). Not included in the initial version: the identity of the direct supervisor. Created by project team from pay transaction files. 1.2. Demographic data. Marital status, gender, and (for those drivers choosing to self-identify to the firm) racial category. Not included, as not captured electronically by firm: educational attainment. Birthdates are automatically included only in 2006 updates and later (see item 3.4). 1.3. Weekly list of drivers employed, with a limited set of driver characteristics. (Used as a check on payroll-based data because payment is not always made in the week activities took place.) 1.4. Weekly list of drivers hired in that week, with a limited set of driver characteristics. 1.5. Weekly list of drivers who separated in that week, with a limited set of driver characteristics. 1.6. Application records. Variables from the application process for those for whom a hire event is observed; has significant number of missing values. Typical items of interest: did applicant request a smoking hotel room when attending the training school, and how did applicant learn about the training opportunity? 2.

Project Component Two: Panel Study of New Hires

2.1. Initial Intake Data on Driver Trainees. Collected from December 12, 2006, through August 8, 2007, on twenty-three Saturdays. 1,069

Using Behavioral Economic Field Experiments at a Firm

101

trainees entered the study, and there is good data on 1,036 of them, except that a bug in the initial version of Hit 15 Points (item 2.1.2.5) limits good data that includes this item to a subset of 893 persons. 2.1.1. First Data Collection Block. 2.1.1.1. Sequential strategic form Prisoner’s Dilemma experiment. Choices as first and second mover, plus estimates of others’ choices are collected. 2.1.1.2. Multidimensional Personality Questionnaire. Standard (short form of) instrument identifying eleven personality factors. 2.1.1.3. Risk/Loss Aversion experiment. Four panels of seven choices, each between a small monetary gamble and a small fixed sum. 2.1.1.4. Demographic profile. List of more than fifty questions compiled by project team, including information on educational attainment, labor market history, next best employment opportunity, and so on. 2.1.1.5. Red Button experiment. Subjects offered opportunity to leave for break early at a payoff cost. 2.1.2. Second Data Collection Block. 2.1.2.1. Time Preferences experiment. Four panels of six choices, each between a smaller amount at an earlier time and a larger amount at a later time. Two panels include “today” as the earlier time, and two panels have a front-end delay, with “Monday” (two days after the Saturday data collection) as the earlier time. 2.1.2.2. Nonverbal IQ: Sections two, three, and four of the original five sections of the Raven’s Progressive Matrices. Participants are asked to rate their relative performance before and after the instrument, as well as whether they want to know their score and the group average upon receiving their final payout. 2.1.2.3. Quantitative Literacy: Section one of Form A of the Quantitative Literacy portion of the Educational Testing Service’s (ETS) Tests of Applied Literacy Skills. (ETS provides standardized scoring only for subjects that do both sections of this test.) Participants are asked to rate their relative performance before and after the instrument, as well as whether they want to know their score and the group average upon receiving their final payout. 2.1.2.4. Ambiguity Aversion experiment. Repetition of Risk/Loss Aversion, except that we provide less information about the probability distribution in the gamble choices. 2.1.2.5. Hit 15 Points backward induction experiment. 2.1.2.6. Miscellaneous questions from the psychology and behavioral economics literature about impatience, risk, cooperation, and so on.

102

Stephen V. Burks et al.

2.2. Follow-up Data from Trainees and their Families. The follow-up data collection period runs through August of 2008. 2.2.1. Human Resources Data from Firm. 2.2.1.1. Subset of Applicant Flow data, including whether trainee requested a smoking room at training. 2.2.1.2. Credit score of the trainee (available because trainees sign a credit contract for training costs). 2.2.1.3. If and when a driver exits, the exit date and a code for up to three reasons for the separation. 2.2.2. Mail Surveys to New-Hire Panel Drivers. 2.2.2.1. Continuing driver surveys. Forty-five questions about current opinions and recent work events, including some openended ones, sent to currently employed drivers between two and four weeks after the following tenure milestones: six months, twelve months, eighteen months, and twenty-four months. 2.2.2.2. Exited driver surveys. Forty-four questions about current opinions, recent work events, and the driver’s exit event, sent between two and four weeks after exit. 2.2.3. Mail Surveys to Families of Trainees. 2.2.3.1. Initial survey to families of new-hire panel drivers, to approximately match in timing the initial intake information from drivers. Not actually started until the beginning of April of 2006, at which time the families of all panel participants who had entered since December 2005, were surveyed. Thirty-five questions developed by project team, including some openended ones. 2.2.3.2. Surveys to families of continuing drivers. Thirty-five questions about current opinions and recent family-work-life events, including some open-ended ones, sent to families of currently employed drivers between two and four weeks after the following tenure milestones: six months, twelve months, eighteen months, and twenty-four months. 2.2.3.3. Surveys to families of exited drivers. Twenty-nine questions about current opinions, recent family-work-life events, and the driver’s exit event, sent between two and four weeks after exit. 2.2.4. Weekly two-question survey to currently employed drivers, in their trucks, via satellite link, on Wednesdays. Question 1: How satisfied are you with your job right now (1–5 Likert response scale)? Question 2: How many pay miles do you expect to run next week? These data will be documented, cleaned, and merged with the main panel study data set, at some point in the future. 2.2.5. Time-at-home records. The firm’s recording of when drivers request to be at home, and how well the firm does at fulfilling those requests, is accumulating in a data file that starts in early 2006, and

Using Behavioral Economic Field Experiments at a Firm

103

is expected to be updated through August 2008. These data will be documented, cleaned, and merged with the main panel study data set at some point in the future. 3.

Control Groups for New Hire Panel Study

3.1. The initial intake data collected on driver trainees (all the items in 2.1) was also collected from 100 undergraduate students recruited at the University of Minnesota, Morris, in a campus computer lab, during the spring semester of 2007. The same protocol was used throughout, with the exception that the demographic profile questions were modified to be appropriate for student subjects. 3.2. The initial intake data collected on driver trainees (all the items in 2.1) is also planned to be collected from 100 adult subjects who are residents of Stevens County, MN, the rural county in which the University of Minnesota, Morris, campus is situated. The plan is to recruit subjects from the county who are not students, nor faculty, nor professional staff, during the summer and fall semesters of 2007. The same protocol will be used throughout, with the exception that the demographic profile questions will be modified to be appropriate for these subjects. 4.

Miscellaneous Supplementary Data

4.1. Records of use of the satellite link in each tractor to report trailers needing maintenance. Available on all drivers by week from mid2006; to be updated periodically. May be usable to generate an index of on-the-job cooperation with other drivers.57 4.2. Records of the use of the satellite link in each tractor to respond to a particular message from the firm’s central dispatch office. Available on all drivers by week from mid-2006, to be updated periodically. May be usable to generate an index of on-the-job cooperation with the firm.58 4.3. Identification of direct supervisors by driver and work week. This 57. When a driver is directed to hook onto a trailer spotted at a customer location but the trailer turns out to be mechanically faulty and can’t be used, it is a negative job event. This happens when the driver who previously had the trailer did not report it as bad, either because he did not realize it was or because he refused to take the time to do so. So when a driver does take the time to report as mechanically faulty a trailer he is dropping and will not immediately use further, thereby delaying the start of his next assignment, he can be said to incur a personal cost to provide a public good for drivers. The main issue in constructing an index is limiting confounds with unobservable factors, such as whether the driver is reporting the trailer as bad because he needs it repaired for his own immediate use. 58. Responding to the macro from the central office is a formal, but not a substantive, job requirement. Responding primarily helps anonymous staff members at the firm’s central office, so if a driver delays proceeding with his current assignment to respond, it can be con-

104

Stephen V. Burks et al.

data is potentially available in a supplementary file that may permit the historical data set used for the statistical case study to be updated. Updates from 2006 and later will include this information in data files already being amalgamated to make the statistical case study data set. 4.4. Birthdates. These are in the standard demographic data files on all drivers starting in 2006. Birthdates for drivers observed in earlier years reside in a supplementary file that must be documented and cleaned and merged with the 2002 to 2005 statistical case study data set. 5.

Merging New Hire Panel and Statistical Case Study Data.

Beginning in the second half of 2008, when the statistical case study data files have been updated to include all of 2007 and at least some of 2008, it is intended to merge the two data sets together for those drivers who are in both data files. This will permit the regression models developed for the statistical case study to be run on the panel study drivers, with the ability to test what additional predictive power is added to those models by the new data items collected in the new hire panel portion of the study.

References Alfred P. Sloan Foundation. 2007. Industry studies. Programs: Standard of living and economic performance. http://www.sloan.org/programs/pg_standard.shtml. ATA Economic and Statistics Group. 2005. Employee turnover and workforce changes: Q4:2004. Trucking Activity Report 13:7. Bandiera, O., I. Barankay, and I. Rasul. 2005. Social preferences and the response to incentives: Evidence from personnel data. Quarterly Journal of Economics 120 (3): 917–62. ———. 2006. The evolution of cooperative norms: Evidence from a natural field experiment. B.E. Journals in Economic Analysis and Policy: Advances in Economic Analysis and Policy 6 (2): 1–26. Beadle, A. D. 2004. Building driver shortage. Traffic World 268 (42): 26–27. Belman, D. L. and K. Monaco. 2001. The effects of deregulation, de-unionization, technology, and human capital on the work and work lives of truck drivers. Industrial and Labor Relations Review 54 (2A): 502–24. Belman, D. L., K. Monaco, and T. Brooks. 2005. Sailors of the concrete sea: A portrait of truck drivers’ work and lives. Lansing, MI: Michigan State University Press. Belzer, M. H. 1995. Collective bargaining after deregulation: Do the teamsters still count? Industrial and Labor Relations Review 48 (4): 636–55. Belzer, M. H., D. Rodriguez, and S. Sedo. 2002. Paying for safety: An economic sidered a personally costly form of cooperation with management. The main issue in constructing an index is limiting confounds with unobservable factors, such as with variations in the level of exposure to such macro requests.

Using Behavioral Economic Field Experiments at a Firm

105

analysis of the effect of compensation on truck driver safety. FMCSA Report. Washington, DC: Federal Motor Carrier Safety Administration. Boyer, K., and S. V. Burks. 2007. Stuck in the slow lane: Traffic composition and the measurement of labor productivity in the U.S. trucking industry. IZA Discussion Paper no. 2576. Bonn, Germany: Institute for the Study of Labor. Bureau of Labor Statistics. 2002. Historical hours and earnings—Average hours and earnings of production or non-supervisory workers on private non-farm payrolls by major industry sector, 1964 to date. U.S. Department of Labor. ftp:// 146.142.4.23/pub/suppl/empsit.ceseeb2.txt. Burks, S. V. 1999. The origins of parallel segmented labor and product markets: A reciprocity-based agency model with an application to motor freight. PhD diss., University of Massachusetts, Amherst. Burks, S. V., J. Carpenter, and L. Götte. 2006. Performance pay and the erosion of worker cooperation: Field experimental evidence. IZA Discussion Paper no. 2013. Bonn, Germany: Institute for the Study of Labor. Burks, S. V., K. Monaco, and J. Myers-Kuykindall. 2004a. The balance between private and for-hire carriage and trends in the use of large trucks (1977 to 1997). Journal of the Transportation Research Forum 43 (2): 159–71. ———. 2004b. Trends in the use of large trucks by truckload and less-thantruckload motor carriers in the 1990s. Journal of the Transportation Research Forum 43 (2): 173–87. Cain, G. G. 1976. The challenge of segmented labor market theories to orthodox theory: A survey. Journal of Economic Literature 11 (4): 1215–57. Camerer, C., and E. Fehr. 2004. Measuring social norms and preferences using experimental games: A guide for social scientists. In Foundations of human sociality: Economic experiments and ethnographic evidence from fifteen small-scale societies, ed. J. Henrich, R. Boyd, S. Bowles, C. Camerer, E. Fehr, and H. Grintis, 55–95. Oxford: Oxford University Press. Casey, J. F. 1987. An assessment of the truck driver shortage. Transportation Executive Update 1 (1): 20–23. Cleves, M. A., W. A. Gould, and R. G. Gutierrez. 2004. An introduction to survival analysis using Stata, revised edition. College Station, TX: Stata Press. Corsi, T., and J. R. Stowers. 1991. Effects of a deregulated environment on motor carriers: A systematic multi-segment analysis. Transportation Journal 30 (3): 4–28. Cox, K. 2004. TL driver turnover rate reaches new record high. Transport Topics (January 5):5, 35. Cullen, D. 1996. You get what you pay for: J. B. Hunt Transport Services Inc.’s driver compensation package. Fleet Owner 91 (11): 82–83. Dickens, W. T., and K. Lang. 1993. Labor market segmentation theory: Reconsidering the evidence. In Labor economics: Problems in analyzing labor markets, ed. W. Darity, Jr., 141–80. Norwell, MA: Kluwer Academic. Fischbacher, U. 2007. z-Tree: Zurich Toolbox for Ready-made Economic Experiments. Experimental Economics, forthcoming. Gallup Organization. 1997. Empty seats and musical chairs: Critical success factors in truck driver retention. Arlington, VA: ATA Foundation. Global Insight, Inc. 2005. The U.S. truck driver shortage: Analysis and forecasts. Arlington, VA: American Trucking Associations. Griffin, G., and L. Kalnbach. 2002. Predicting and classifying voluntary turnover decisions for truckload drivers. Report no. MPC-02-135. Fargo, ND: Upper Great Plains Transportation Institute, North Dakota State University. Griffin, G. C., J. M. Rodriguez, and B. M. Lantz. 1992. Job satisfaction in U.S.

106

Stephen V. Burks et al.

commercial drivers. Report no. DP-90. Fargo, ND: The Upper Great Plains Transportation Institute, North Dakota State University. Harrison, G. W., and J. A. List. 2004. Field experiments. Journal of Economic Literature 42 (4): 1009–55. Isidore, C. 1997. Hunt gambles on pay raise to keep drivers. Journal of Commerce 411 (28917): 1A. Kiefer, N. M. 1988. Economic duration data and hazard functions. Journal of Economic Literature 26:646–79. Mishel, L., J. Bernstein, and S. Allegretto. 2006. The state of working America, 2006–2007. Ithaca, NY: ILR Press. Patrick, C. J., J. J. Curtin, and A. Tellegan. 2002. Development and validation of a brief form of the Multidimensional Personality Questionnaire. Psychological Assessment 14 (2): 150–63. Raven, J., J. C. Raven, and J. H. Court. 2000. Section 3: The Standard Progressive Matrices. In Manual for Raven’s Progressive Matrices and Vocabulary Scales. San Antonio, TX: Psychological Corporation. Reiskin, J. 2006. Shipper changes cut distances that truckload freight travels. Transport Topics (March 13):1, 49. Rodriguez, J., M. Kosir, B. Lantz, G. Griffin, and J. Glatt. 2000. The costs of truckload driver turnover. Report no. SP-146. Fargo, ND: Upper Great Plains Transportation Institute, North Dakota State University. 1–18. Stephenson, F. J., and R. J. Fox. 1996. Driver retention solutions: Strategies for forhire truckload (TL) employee drivers. Transportation Journal 35 (4): 12–26. U.S. Census Bureau. 2004. Truck transportation: 2002. Washington, DC: Economics and Statistics Administration, U.S. Department of Commerce. Waxler, C. 1997. J. B. Hunt Transport shares may drop further. Forbes 159 (8): 382–83.

3 Subjective Evaluation of Performance and Evaluation Interview Empirical Evidence from France Marc-Arthur Diaye, Nathalie Greenan, and Michal W. Urdanivia

“If the economic organization meters poorly, with rewards and production only loosely correlated, then productivity will be smaller; but if the economic organization meters well productivity will be greater.” —Alchian and Demsetz (1972, 779)

3.1 Introduction A main contribution of linked longitudinal employer-employee data is to provide a decomposition of wage rates into components due to individual heterogeneity and to firm heterogeneity. In France, Abowd, Creecy, and Kramarz (2002) show that the person effect and firm effect account, respectively, for 70 percent and 20 percent of the variation of wages. The person-effect component is bigger in France than in the United States where it represents half of the wage variation. This indicates that the devices used by firms to attract or select workers with specific characteristics play a central role in determining the firm’s wage structure. However, these devices have not been investigated thoroughly by economic analysis. In this paper, we are going to assess individual evaluation Marc-Arthur Diaye is an associate professor at the Center for the Study Political Economics at the University of Evry, and a research fellow at the Center for Labor Studies. Nathalie Greenan is a researcher at the National Centre for Scientific Research and a research director at the Center for Labor Studies. Michal W. Urdanivia is a PhD student in the Department of Applied Mathematics at the University Panthéon-Sorbonne, and a research assistant with the Center for Labor Studies. This paper was presented at the Conference on Analysis of Firms and Employees (CAFE 2006) held September 29 to 30, 2006, in Nuremberg, Germany. We gratefully acknowledge the financial support provided by the Institute for Employment Research (IAB), the Data Access Center (FDZ-BA/IAB), The Deutsche Forschungsgemeinschaft (German Research Foundation), their Research Network “Flexibility in Heterogeneous Labour Markets,” the Alfred P. Sloan Foundation, and the National Science Foundation. We are deeply indebted to Emmanuel Duguet, Lynn Zucker, and the two referees who helped us to improve this paper. We thank the participants at the CAFE 2006, the BETA-Cereq (University Louis Pasteur) seminar on labor markets, the PSE-Jourdan seminar on labor markets and inequalities (especially Andrew Clark), and the TEAM Seminar of the Centre d’Economie de la Sorbonne for their comments.

107

108

Marc-Arthur Diaye, Nathalie Greenan, and Michal W. Urdanivia

interview, a human resource management (HRM) practice that could contribute to the two goals of selecting workers and stimulating their effort. In France, 52 percent of employees with more than one year of seniority in manufacturing had been evaluated at least once in 1997. At that time, evaluation interviews were not regulated at the national or at the industry level. As an HRM practice, the function of evaluation interviews is not clearcut. Sometimes viewed as formal performance appraisal systems, evaluation interviews often use complex evaluation grids referring to loosely defined behavioral characteristics as well as to precisely defined goals and measured criteria. To assess evaluation interviews, it is useful to analyze them theoretically and to investigate empirically how they are implemented within firms. The theoretical framework we are going to use in this paper is the one (hereafter the DGU model) proposed by Diaye, Greenan, and Urdanivia (2007). Intuitively, individual evaluation interviews are used to assess performance once the employee has undertaken her or his task. We will use here the term of ex post evaluation interviews. But, if we refer to a classic wagesetting mechanism, there is no need for evaluation. The incentive wages drives the employee toward the level of effort that is optimal for the employer. In the DGU model, ex post evaluation interviews insure risk adverse agent against technological or market uncertainty. In the French context, Crifo, Diaye, and Greenan (2004) observe that evaluation interviews are significantly more frequent when the employee is involved in collective work. The model of ex post evaluation interviews is limited to give account of the function of evaluation interviews when employees work in teams because individual performance appraisal become difficult when the output cannot be separated between the members of a team (Alchian and Demsetz 1972). Diaye, Greenan, and Urdanivia (2007) propose a model of ex ante individual evaluation interviews specific to the teamwork context, where evaluation comes first, before the constitution of teams and aims at fostering a team spirit. They conclude their theoretical approach by establishing some predictions about drivers and outcomes of individual evaluation interviews. In this paper, we want to assess empirically part of these predictions. Our empirical investigation rests on a matched employer-employee survey (section 3.2) on organizational change and information and communication technology (ICT) use (computerization and organizational change [COI]). In the labor force section of the survey, employees are asked whether they have been interviewed individually at least once in 1997. They also give information on work organization, on personal characteristics, and on outcomes. The business section of the survey gives a detailed set of firm-level characteristics reflecting technological and organizational choices implemented in French manufacturing at the end of the 1990s. We use a propen-

Subjective Evaluation of a Performance and Evaluation Review

109

sity score methodology (section 3.3) to evaluate (section 3.4) the causal effect of individual evaluation interviews on effort, work overload, and wage setting. In section 3.5, we conclude. 3.2 The Data We are going to use a matched employer-employee survey, the survey on computerization and organizational change (COI), to assess the DGU model of evaluation interviews. The information we have in the survey will not allow us to test all the predictions from their model. However, from the employee section of the survey, we have some information on the characteristics of work (whether individual or collective), on evaluation interviews, on effort, and on wages. This will allow us to cover the main features underlined by the DGU model. More precisely, we will be able to test whether evaluation interviews lead to higher levels of effort than classical incentive schemes (prediction 1). Our estimation strategy will also allow us to assess the existence of a selection effect associated with the implementation of evaluation interviews in individual and collective work organizations (prediction 2). Furthermore, using measures of work overload, we will check whether evaluation interviews drive workers toward an excessive work intensity leading to inefficiencies (prediction 3). Indeed, according to the DGU model, evaluation interviews in a context of supermodular technology (i.e., the conditional probability of success of the task is a strictly increasing convex function of the employees’ level of effort) lead to an overintensification of work in the sense that the employees’ level of effort will be higher than the one “required” by the firm. The reason is the selection effect regarding disutility of effort. Prediction 3 is a possible consequence of this result on overintensification. It is important to test this implication because work overload is a major factor of stress and has longterm implications on the health of the workforce, especially in a context of aging. Finally, we will be able to test our predictions on wage differentials and on the employees’ knowledge of the rules driving wage setting between the scheme with evaluation interviews and the classical incentive scheme (prediction 4). The COI survey was conducted at the end of 1997 by the French public statistical system.1 We are going to work on a representative sample of 1. The conception and coordination of the COI survey has been directed by the Center for Labor Studies. The survey has been carried out in a consortium involving the Ministry of Labor (DARES), the Ministry of Industry (SESSI), the Ministry of Agriculture (SCEES), and the National Institute of Statistics and Economic Studies (INSEE). It benefited from very high response rates: 82 percent for employers and 75 percent for the employees. For a detailed description of the survey, see Greenan and Hamon-Cholet (2001) or http://www .enquetecoi.net.

110

Marc-Arthur Diaye, Nathalie Greenan, and Michal W. Urdanivia

manufacturing firms with more than fifty employees and on a sample of randomly selected employees within these firms. In matched employeremployee surveys, the budget constraint implies a trade-off between trying to capture the diversity of firms and trying to capture the diversity of the workforce within firms. By choosing to interview small sample of employees (one, two, or three) within each firm, COI chooses to favor the diversity of firms. As interviewed employees have at least one year of seniority within the firm, they belong to its core workforce. In the full sample of the labor force section of the survey, there are 4,295 employees. However, in our analysis, we do not take into account employees with supervision activities (1,214 individuals) or employees working part time (177 individuals). Indeed, the former combine a position of Principal and of Agent that we have not investigated theoretically, while part time leads to badly measured effort and wages. We obtain a subsample of 2,904 employees. The available information on the practice of individual evaluation interviews stems from the following question: Do you have at least one evaluation interview per year (yes / no)? Because of their seniority in the firm, we know that all interviewed employees had the opportunity of being evaluated at least once. The labor force section of the COI survey describes in detail work organization. It includes a whole set of questions capturing whether work is structured around group activities. From these questions, we build up five different measures of interaction between employees in the course of the work process: being part of a team, time spent in teamwork, intensity of communication with other workers, level of support from other workers, participation in meetings (see appendix A for detailed questions). These five measure are positively correlated, with correlations ranging between 0.04 (intensity of communication with time spent in team work) and 0.18 (being part of a team and level of support from other workers). Thus, they measure different dimensions of collective work. We derive from these five measures a synthetic binary indicator of collective work. When it takes the value 1, the employee is considered as being a “collective” worker, when it takes the value 0, he or she is considered as being an individual worker. According to this variable, our sample of employees breaks down into 1,537 individual workers and 1,367 collective workers. Table 3.1 gives the distribution of individual evaluation interviews according to our synthetic binary indicator of collective work. In 1997, 37.2% of the employees have been interviewed at least once. Evaluation interviews are positively correlated with collective work: 47 percent of collective workers have been evaluated against 29 percent of individual employees. The COI survey also measures different effort indicators. Productive effort is measured through two questions indicating if the employee works

Subjective Evaluation of a Performance and Evaluation Review Table 3.1 Evaluation Yes No Total a

111

Evaluation among individual workers and collective workers Individual workers

Collective workers

445 (29%) 1,092 (71%)a 1,537 (53%)

637 (47%) 730 (53%) 1,367 (47%)

Percentage with respect to the analyzed subsample of 2,904 employees.

longer than the usual hours some days or some weeks. Productive effort is considered as very high if the employee sometimes increases hours worked for personal reasons, as high if he or she sometimes increases hours worked in response to the firm’s demand, and as low if longer hours never happen. According to these three situations, the productive effort indicator, respectively, takes a value of 2, 1, or 0. The cognitive effort indicator is a binary variable indicating if the employee makes propositions to improve his or her workstations, the production process, or the machines. It measures an involvement in collective knowledge building about the productive activity, allowing continuous improvement of the production process. Two additional measures are included in the analysis to identify if effort is going beyond reasonable levels, creating an overload that could be detrimental for work efficiency and for the employee’s health. A first variable indicates how often an employee has to hurry in the course of his or her work. Four states are taken into account: hurrying almost all the time, hurrying for one quarter of the time or more, hurrying for less than a quarter of the time, and never. The hurry variable, respectively, takes the value 4, 3, 2, and 1 according to the intensity of the pressure. Work overload is also measured through a binary indicator telling whether the employee often has to interrupt one task to carry out another urgent and nonanticipated one. Finally, we measure the employee’s annualized net wage in euros. As it comes from an administrative data file used to compute social contributions, it is precisely measured and includes all bonuses, taxed allowances, and compensations in kind. We also build up an indicator of the employees’ ability to predict their wages. It rests on a question about the elements that have a big influence on the employee’s wage or promotion, followed by a list of eight items. We compute the ratio of the number of yes responses to the list of items, on the number of yes and no, which gives an indicator taking its value between 0 and 1. Zero means that the employee has no idea of how to increase his or her wage or chance of promotion, 1 means that the employee knows that he or she can improve his or her situation and is aware of what to do to obtain this outcome.

112

Marc-Arthur Diaye, Nathalie Greenan, and Michal W. Urdanivia

3.3 Estimation Strategy We want to measure the impact of evaluation interviews on effort, work overload, and wages, but we know from the DGU model that evaluation interviews induce a selection process. Employees with a low disutility of effort and, in the case on teamwork, with a team spirit are going to be attracted by jobs where evaluation interviews are conducted periodically. A possible way to measure outcomes related to evaluation interviews, taking into account the selection effect, is to consider evaluation interviews as treatments and to apply a propensity score method to match each treated individual with a nontreated individual with the same characteristics in order to turn our nonexperimental data into a quasi experiment. A simple way to test the predictions of the DGU model is to consider evaluation interviews as treatments and to evaluate the effect of this treatment on the chosen variables for measuring effort, wages, and beliefs about wages. More precisely, let t be a dummy variable equal to 1 if the employee declares being evaluated and 0 otherwise. Three quantities are of interest to us. The first is the average treatment effect over the whole population, written C; the second is the average treatment effect over the treated individuals, written C1; and the third is the average treatment effect over the nontreated individuals, written C0. More precisely, let Y be the chosen variables for measuring effort, wages, and beliefs about wages. Then C measures the variation of Y that would be observed if the whole population was treated; C1 is an evaluation of the effect of the treatment in the usual sense because it concerns the treated population; and C0 is a prospective evaluation in the sense that it measures what would happen if the nontreated population was treated. We have: C  E(Y1 – Y0) C1  E(Y1 – Y0 | t  1) C0  E(Y1 – Y0 | t  0), where Y1 is the observed value of Y that results when receiving treatment (that is, when being evaluated), Y0 is the observed value of Y that results when not receiving treatment (that is, when not being evaluated), and E(.) denotes expectation in the population. Intuitively, an estimate of an average treatment effect could be the difference between the average of Y over the population of treated individuals and its average over the population of nontreated individuals, that is, Y 1 – Y 0, where Y 1 and Y 0 are, respectively, the average of Y for treated (evaluated employees) and the nontreated (nonevaluated employees).

Subjective Evaluation of a Performance and Evaluation Review

113

However, broadly speaking, the main problem when evaluating the effect of a treatment is that for each individual we only observe Y  t  Y1  (1 – t)  Y0 



Y1 if t  1 . Y0 if t  0

Then it can be shown that the average difference between treated and nontreated individuals can be the cause of a selection bias because the data does not result from a randomized experiment. And when testing evaluation effects (on effort, overload, and wages), there is a need to control for naturally occurring systematic differences in background characteristics between the treated population and the nontreated population, which would not occur in the context of a randomized experiment. Moreover, according to prediction 2, individual evaluation interviews affect employees’ efforts through a selection effect associated to disutility or to team spirit, an incentive effect that in our case is estimated by the average treatment (evaluation) effect. Therefore, in order to estimate the average treatment (evaluation) effect, it is also necessary to control for the selection bias due to disutility. Although it seems difficult to control “directly” for this selection effect because disutility or team spirit are not observable characteristics, we can assume that they are grounded on observable background characteristics of the employee and of the employer, and, hence, controlling for them allows to control for the selection. We will discuss in the next section the background characteristics we will take into account to estimate the effect of individual evaluation interviews. We choose to use the propensity score methodology introduced by Rosenbaum and Rubin (1983). This method reduces the entire collection of background characteristics to a single composite characteristic that appropriately summarizes the collection. Propensity score technology allows to correct the selection bias by matching individuals according to their propensity score, which is the estimated probability of receiving the treatment (of being evaluated) given background characteristics. We are going to use a nonparametric kernel matching estimator proposed by Heckman, Ichimura, and Todd (1997, 1998), which under some regularity assumptions is convergent and asymptotically normal. 3.4 The Results 3.4.1 Determinants of Individual Evaluation Interviews The first step of the propensity score method is to analyze the determinants of evaluation interviews, taking into account background characteristics that influence the employee’s probability of receiving a periodical

114

Marc-Arthur Diaye, Nathalie Greenan, and Michal W. Urdanivia

evaluation interview and the three categories of outcomes we consider: effort, work overload, and wages. In this step, it is very important to take into account individual effects as well as contextual effects. As we have pointed out, personal characteristics of the employee like team spirit or disutility of effort are going to play a crucial role in influencing both the chances of being evaluated and the outcomes we consider. These characteristics are not directly observable, but we are going to take into account observables that are possibly correlated with them: gender, age, seniority, education level, and occupation. It is clear that these personal characteristics have impacts on effort levels, work overload, and wages. The fact that our employee sample is matched with a survey describing the characteristics of firms is an important advantage in our estimation strategy. The DGU model has stressed that the production technology plays a role in the diffusion of evaluation interviews. A supermodular technology is more favorable than a submodular technology. In order to control for the technology, we are going to include the regression size and sector dummies. Stemming from an employer database, information on size and sector is much more precise than the information usually included in labor force surveys. We also include a measure of the firm’s computerization intensity. We choose to build up a variable describing the intensity of numerical data transfers within and outside the firm. Moreover, evaluation interviews could be complementary to other organizational practices, and these practices could also have an influence on outcomes. Eight new organizational practices are considered in the logistic regression: quality certification, total quality management, methods to analyze products and processes (value analysis; functional analysis; Failure Mode, Effects, and Criticality Analysis [FMECA]), total productive maintenance (TPM), organization in profit center, formal in-house customer/supplier contracts, system of justin-time delivery, and system of just-in-time production. We also detail different teamwork practices: self-managed teams, problem solving groups, and project teams. Finally, we take into account the evolution of the number of hierarchical layers in the firm and variables indicating difficulties connected with the implementation of organizational changes. Appendix C presents the parameters estimated of the logistic models explaining individual evaluation interviews for individual workers and for collective workers. In the case of individual workers, we find that employee characteristics have higher explanatory power than employer characteristics. More precisely, male workers in executive or middle management positions with either low seniority (one or two years) or intermediate seniority (seven to ten years) have a higher probability of being evaluated. We have to keep in mind that even though some of the interviewed workers have management positions, they have no formal hierarchical authority as they declare no subordinates. Among the employer characteristics, the

Subjective Evaluation of a Performance and Evaluation Review

115

only variables with significant influence are size, with a positive impact of the highest size cluster; sector, with a positive impact of five sectors (pharmaceutical, perfumes, and cleaning products; chemicals, rubber, and plastic products; electrical and electronic equipment; electrical and electronic components; and shipbuilding, aircraft, and railway); and quality certification (ISO 9001, ISO 9002, and EAQF). In contrast, in the case of collective workers, employer characteristics tend to explain more than employee characteristics. Indeed, for team workers the only personal characteristic that influences the probability of being evaluated is the level of education: a second or third level of education is associated with a coefficient that is positive and significant. On the employer side, size, sector, computer intensity, use of new organizational devices, and use of teamwork have a significant impact on the probability of being evaluated. Employers with medium size (between 100 and 999 employees) and belonging to pharmaceutical, perfumes, and cleaning products or to chemicals, rubber, and plastic products use evaluation interviews more frequently. Employers from printing, press, and publishing and shipbuilding, aircraft, and railways have a lower probability of being interviewed. The intensity of computerization favors evaluation interviews of collective workers as well as quality certification and total productive maintenance. Conversely, employers using just-in-time delivery are less oriented toward evaluation interviews for collective workers. Having a nonmarginal fraction of production workers in problem solving groups favors evaluation interviews, while having a small fraction of nonproduction workers participating in self-managed teams and having management involved in project teams has a negative impact on evaluation interviews. In total, evaluation interviews for collective workers seem complementary with information technologies and new organizational practices. These managerial tools could support a supermodular production technology, where the employer has a preference for higher levels of effort. 3.4.2 Observing the Outcomes of Individual Evaluation We are now going to discuss the matching evaluation of the effect of individual evaluation interviews on individual and collective workers on effort (table 3.2), work overload (table 3.3), and wages (table 3.4). In each table, we first compute as a benchmark the average outcome for individual and collective workers. Second, we compute the average difference in outcome between workers that have been individually evaluated and workers that have not been evaluated. This estimator is often designated as the naive estimator of the treatment effect. Then we compute the three causal effects: the effect on the treated (C1), the effect on the nontreated (C0), and the global effect (C). The first effect is the matching evaluation strictly speaking, the second one represents the effect that evaluation interviews would have if they were implemented on the nonevaluated population of

116

Marc-Arthur Diaye, Nathalie Greenan, and Michal W. Urdanivia

workers, and the last one is the effect that would be obtained if evaluation interviews were extended to the entire population. Effort We observe higher levels of productive and cognitive efforts when work is collective rather than individual (table 3.2). This was not entirely expected because our model underlined that one of the advantage of collective work was to share the burden of higher levels of effort between workers. However, other effects might play a role here. The DGU model (as well as the analysis of determinants of evaluation interviews) suggests that collective work is positively correlated with supermodular production technologies. Another explanation could lie in synergy and peer pressure effects connected with collective work. As predicted by the DGU model, we observe that the level of effort, whether productive or cognitive, is higher when workers are individually evaluated than in the classical incentive scheme (prediction 1). The causal treatment effect on productive effort is stronger for individual workers than for collective workers. And the selection effect has an opposite sign. Individual workers displaying higher level of effort are selected in the population of evaluated workers, when they are selected out in the Table 3.2

Individual evaluation interviews and effort

Productive effortc Average productive effort Average difference E/NE Effect on the treated (C1) Effect on the nontreated (Co) Global effect (C) Cognitive effort c Average cognitive effort Average difference E/NE Effect on the treated (C1) Effect on the nontreated (Co) Global effect (C)

Individual workersa

Collective workersb

0.564 0.127*** 0.084** 0.093** 0.091**

0.720 0.092*** 0.120** 0.100** 0.110**

0.507 0.143*** 0.099** 0.120** 0.114**

0.722 0.140*** 0.110** 0.110** 0.110**

a The standard deviation of the treatment effect is computed using bootstrap with 300 simulations. The characteristics of the support over 300 simulations are min = 1,352; max = 1,501; mean = 1,426.48. b The standard deviation of the treatment effect is computed using bootstrap with 300 simulations. The characteristics of the support over 300 simulations are min = 1,124; max = 1,304; mean = 1,229.03 c See section B of appendix A for a description of these variables. ***p-value < 0.01. **0.01 ≥ p-value < 0.05. *0.05 ≥ p-value < 0.1.

Subjective Evaluation of a Performance and Evaluation Review

117

case of collective work. This result corroborates prediction 2 although the DGU model gives no specific clue to understand our surprising result on collective workers. The extension of evaluation interviews to the whole population of collective workers would consequently increase productive effort although it is already high in this case. The observed effects on cognitive effort are more straightforward. Evaluation interviews similarly affect cognitive effort for individual and collective workers: they increase by 14 percent the propensity to make propositions for improving the production process. In the case of cognitive effort, the selection effect has an identic sign among individual and collective workers, but it is stronger in the first case. Work Overload Individual and collective workers work with a similar time pressure: the average need to hurry is 2.67 in the first case, 2.64 in the second (table 3.3), indicating that workers have to hurry a little more than a quarter of their time. Our second indicator of work overload is higher for collective workers: 65 percent of collective workers experience task interruptions in the course of their work, whereas 53 percent of individual workers face interruptions. However, it is in the case of individual workers that evaluation interviews have a significant impact as it appears to mitigate work overload. Individual workers that are periodically evaluated work under lower time pressure

Table 3.3

Individual evaluation interviews and work overload

Hurry c Average overload Average difference E/NE Effect on the treated (C1) Effect on the nontreated (Co) Global effect (C) Interruptc Average overload Average difference E/NE Effect on the treated (C1) Effect on the nontreated (Co) Global effect (C) a

Individual workers a

Collective workersb

2.666 –0.143** –0.142 (ns) –0.189** –0.176**

2.640 –0.110* –0.108 (ns) –0.073 (ns) –0.089 (ns)

0.526 –0.053* –0.065** –0.066** –0.066**

0.650 0.009 (ns) 0.002 (ns) –0.003 (ns) –0.000 (ns)

See table 3.2 footnote. See table 3.2 footnote. c See section C of appendix A for a description of these variables. ***p-value < 0.01. **0.01 ≥ p-value < 0.05. *0.05 ≥ p-value < 0.1. b

118

Marc-Arthur Diaye, Nathalie Greenan, and Michal W. Urdanivia

and are less exposed to task interruptions. In the case of time pressure, the selection effect seems to play an important role as the causal effect on the treated is not significant. But evaluation interviews also seem to have a protective effect on their own because the effect on the nontreated is negative, significant, and stronger than the naive estimator. Individual workers who have been selected out from evaluation interviews would benefit from their implementation. As far as task interruptions are concerned, the protective effect of evaluation interviews is not explained by a selection effect; it is a pure outcome of this managerial device. Evaluation interviews do not protect collective workers from work overload, but they do not increase their risk of exposition either. It is also an interesting result, knowing that collective workers produce higher levels of productive and cognitive efforts. These results could be evidence of prediction 3. Evaluation interviews in a context of supermodular technology lead to an overintensification of work, but not to work overload. On the contrary, they seem to mitigate work overload, either through a selection effect as described in the DGU model, or through a pure effect. Wage Setting Collective workers earn more, on average, than individual workers (table 3.4). We also observe that, on average, evaluated employees earn more

Table 3.4

Individual evaluation interviews and wage setting

Annualized net wage (in euros)c Average net wage Average difference E/NE Effect on the treated (C1) Effect on the nontreated (Co) Global effect (C) Employee’s ability to predict his or her wagec Average ability to predict Average difference E/NE Effect on the treated (C1) Effect on the nontreated (Co) Global effect (C) a

Individual workers a

Collective workers b

15,003 1,654*** 198 (ns) 275 (ns) 253 (ns)

16,586 1,925*** 1,310** 1,062** 1,174**

0.491 0.164*** 0.145*** 0.147*** 0.146***

0.597 0.136*** 0.110*** 0.100*** 0.100***

See table 3.2 footnote. See table 3.2 footnote. c See sections D and E of appendix A for a description of these variables. ***p-value < 0.01. **0.01 ≥ p-value < 0.05. *0.05 ≥ p-value < 0.1. b

Subjective Evaluation of a Performance and Evaluation Review

119

than employees in a classical incentive scheme, confirming prediction 4. These monetary gains are higher for collective than for individual workers: 1,925 euros per year, on average, against 1,654 euros per year. For individual workers, this difference is entirely explained by the selection effect: the causal effects on the treated is not significantly different from zero, and the causal effect on the nontreated is also nonsignificant. Contrary to individual workers, the monetary gain of collective workers is only slightly lower when selection is taken into account: the gain falls from 1,925 euros to 1,310 euros if we consider the causal effect on the treated, to 1,062 if we consider the causal effect on the nontreated, and to 1,174 if we consider the global effect. Concerning the employee’s ability to predict his or her wage, we first note that this ability is greater, on average, for collective workers than for individual workers, and in both cases the average difference between evaluated and nonevaluated workers is significantly different from zero. Moreover, this effect of evaluation interview still remains significant when one corrects for the selection effect. As stated by prediction 4, evaluated workers have a better knowledge of the rules driving wage setting. 3.5 Conclusion Diaye, Greenan, and Urdanivia (2007) have proposed a theoretical framework based on a Principal-Agent model to analyze the underlying mechanisms of individual evaluation interviews in the case of individual production and of team production (DGU model). They distinguish an ex post evaluation interview that builds a subjective evaluation of employees’ effort and an ex ante evaluation interview which, in the case of team production, works as a coordination device through the fostering of a team spirit. Their theoretical analysis allows deriving testable predictions regarding the effect of individual evaluation interviews on productive and cognitive effort, on work overload, and on wage setting. Using a matched employer-employee survey on computerization and organizational change (COI), we are able to test part of these predictions and to corroborate them. First, evaluation interviews have a positive impact on productive and cognitive effort. Second, evaluation interviews increase effort through two effects: the classical incentive effect and also a selection effect. Third, the selection effect is stronger in the case of individual production compared with the case of team production. Fourth, evaluated employees earn more than employees in a classical incentive scheme, and fifth, evaluated workers have a better knowledge of the rules driving wage setting. The DGU model also suggests a higher propensity to evaluate workers in firms when the production technology is of a supermodular type and an overintensification of work in such a technological context. Our empirical

120

Marc-Arthur Diaye, Nathalie Greenan, and Michal W. Urdanivia

results tend to indicate that collective work is positively correlated with supermodular technologies as collective workers are more frequently evaluated and provide a higher level of effort than individual workers. However, evaluation interviews are not associated with work overload. On the contrary, individual workers seem to be protected from work overload when they are evaluated, and collective workers do not register a higher exposure to work overload even though they provide higher levels of effort. This could be an indirect evidence of the selection effect already stressed. Evaluated workers produce higher levels of effort, but their personal characteristics or the characteristics of their employers allow them to better cope with it.

Appendix A Variables Constructed from the Labor Force Section of the COI Survey A. Measures of Collective Work In the COI survey, a sample of randomly selected employees within interviewed firms (one, two, or three per firm) are asked to describe in detail the way they work at the time when they are being surveyed. Measure 1: Teamwork This measure is associated to the following question: “Do you sometimes do your work in group or collectively?” Response is either “yes” or “no.” Measure 2: Time Spent in Teamwork This measure is constructed from the following question asked to employees who declared working in group or collectively: “How much of your working time do you work in group or collectively?” Responses are “Almost all the time,” “More than a quarter of your time,” “Less than a quarter of your time.” Measure 3: Communication in the Firm This measure is constructed from the following four questions: “Apart from your superiors, are there other persons who give you indications on what you have to do?” (Responses are either “yes” or “no,” or “it does not apply.”) 1. “Colleagues you usually work with?” 2. “Other persons or departments in the firm?”

Subjective Evaluation of a Performance and Evaluation Review

121

“Apart from your subordinates, do you give indications to other persons on what they have to do?” (Responses are either “yes” or “no” or “it does not apply.”) 3. “Colleagues you usually work with?” 4. “Other persons or departments in the firm?” Then a low intensity of communication corresponds to 0 or 1 positive answer among these four questions, and a high intensity of communication to at least two positive answers among the four questions. Measure 4: Support from Other Workers The measure is constructed from the following three questions: “If you have a temporary excess workload or if you are uneasy with a difficult task, are you helped by . . .” (Responses are either “yes” or “no” or “it does not apply.”) 1. “your superiors?” 2. “colleagues you usually work with?” 3. “other persons or departments in the firm?” Then a low level of support from other workers corresponds to 0 or 1 positive answer, and a high level of support corresponds to at least two positive answers among the three questions. Measure 5: Participation in Meetings The measure is constructed from the following question: “How many times a year do you participate in meetings in the context of your work?” Then a low participation in meetings corresponds to 0 or only one meeting a year, and a high participation in meetings corresponds to at least two meetings a year. (See tables 3A.1 and 3A.2.) B. Measures of Effort Two dimensions of effort are captured in the COI survey. The first one describes the level of productive effort. It is built from the answers to the two following questions: (1) “Do you work more than ordinarily?” (Response is either “yes” or “no”); (2) If yes, “Do you work more than ordinarily for personals reasons?” (Response is either “yes” or “no.”) Employee’s effort is a variable with three levels: 0 when the answer to question (1) is “no,” 1 when the answer to question (1) is “yes” and the answer to question (2) is “no,” 2 when both the answers to questions (1) and (2) are “yes.” The second one describes the level of cognitive effort or, more precisely, the degree of implication into collective knowledge building about the pro-

122

Marc-Arthur Diaye, Nathalie Greenan, and Michal W. Urdanivia

Table 3A.1

Five measures of interaction between employees in the work process Frequency

Percent

1,422 1,482

48.97 51.03

2,045 859

70.42 29.58

1,019 1,885

35.09 64.91

1,537 1,367

52.93 47.07

1,557 1,347

53.62 46.38

Measure 1: Teamwork No (0) Yes (1) Measure 2: Teamwork intensity Less than 1/4 of time (0) 1/4 of time or more (1) Measure 3: Communication intensity Low (0) High (1) Measure 4: Level of support Low (0) High (1) Measure 5: Participation in meetings Low (0) High (1)

Table 3A.2

Measure 1 Measure 2 Measure 3 Measure 4 Measure 5

Correlation coefficients between the five measures of interaction between employees Measure 1

Measure 2

Measure 3

Measure 4

Measure 5

1.00000 0.63486 0.15159 0.17993 0.13480

1.00000 0.04493 0.11736 0.04775

1.00000 0.14987 0.17168

1.00000 0.08567

1.00000

duction process. It is built from the answers to the following questions: (1) “In the context of your work, do you make propositions to improve your workstation, the production process, the machines . . . ?” (Response is either “yes” or “no.”) C. Measures of Work Overload Two indicators measure work overload: • Hurry is a discrete variable. It is equal to 1 if the employee states that he or she never has to hurry to do his work. It is equal to 2 if he or she states that it is the case for less than one quarter of the time. It is equal to 3 if he or she states that it is the case for one quarter of the time or more. And it is equal to 4 if he or she states that he or she has to hurry almost all the time. • Interrupt is a dummy variable equal to 1 when the employee states that

Subjective Evaluation of a Performance and Evaluation Review

123

he or she often has interrupt one task to carry out another urgent and nonanticipated one. It is equal to 0 otherwise. D. Measure of Monetary Incentives Monetary incentives are captured through annualized net wage (in euros). It comes from the annual declarations of social data (DADS), which is an administrative file used to compute the tax on wages. It groups all earnings paid in cash or kind between the 1st of January and the 31st of December 1996, less social contributions (social security, pensions, and unemployment benefit). This compensation includes base wage, all bonuses, taxed allowances, and compensations in kind. Bonuses associated with the two French profit sharing regimes (participation and intéressement) are not included when they are not taxed. However, bonuses connected to participation schemes are generally not taxed when the reverse is true for bonuses connected to intéressement schemes. It is the length of the period during which bonuses remain unavailable that determines taxation. In the case of participation, when this unavailability period is shortened to three years, the bonuses become partly eligible to taxation. In the case of intéressement, bonuses are partly exonerated from taxes when they are blocked for a while in a company saving scheme. If we except bonuses connected with participation, compulsory in firms with more than fifty employees, most of the earnings that contribute to an individualization of compensations are taken into account in our variable. Thus, we may interpret it as an output of the wage policy of the firm. Last, compensations correspond to employment periods that vary from one employee to the other. We have annualized the information we had, taking into account the number of days worked. This does not correct for part time, but only 6 percent of the employees in our sample declare working part time. E. Employee’s Ability to Predict His or Her Wage This variable is built from the answers to the eight following questions: “Which of the following elements have a big influence on your wage or on your promotion” (for each element, response is either “yes” or “no” or “it does not apply”): (1) “To do a high-quality work?”; (2) “To carry assignments to the letter?”; (3) “To be on good terms with the boss (bosses)?”; (4) “To be on good terms with the colleague(s)?”; (5) “To take up training courses?”; (6) “To learn how to use new technologies?”; (7) “The firm’s performances?”; (8) “Other reasons?” The employee’s ability to predict his or her wage is then the ratio of number of “yes” answers to the number of “yes” or “no.”

124

Marc-Arthur Diaye, Nathalie Greenan, and Michal W. Urdanivia

Appendix B Variables Constructed from the Firm Section of the COI Survey A. Firms’ Computerization Intensity This variable in constructed from the following question: “Did/Does your firm realize data transfers by means of a computer interface . . .” (Response is either “yes” or “no”): 1. “within the management service?” 2. “between management and production service?” 3. “between management and suppliers, subcontractors?” 4. “between management and client firms?” 5. “between management and social organisms public power?” 6. “between conception services and production?” 7. “between conception and suppliers, subcontractors?” 8. “within the production services or between manufacturer unities?” 9. “between production and suppliers, subcontractors?” 10. “between production and client firms?” Computerization intensity is equal to 1 if there is 0 or 1 “yes”; intensity 2 corresponds to two or three “yes”; intensity 3 corresponds to four or five “yes”; and intensity 4 corresponds to five and more “yes.” B. Average Number of Tasks That Each Type of Worker Is Responsible for (NMT) This variable is constructed from the question (responses are “Management”/“Production Worker”/“Specialist”: more than one answer is possible for each subquestion). “In general, who is/was authorized in 1997 to . . .”: 1. adjust installations? 2. perform first level maintenance? 3. allocate tasks to production workers? 4. inspect quality of supplies? 5. inspect quality of production? 6. participate in performance improvements? 7. participate in project teams? 8. stop production in case of an incident? 9. troubleshoot in case of an incident? 10. start production again in case of an incident? The qualitative variable NMT with four items is constructed as follows:

Subjective Evaluation of a Performance and Evaluation Review

125

NMT  1.7 (High responsibility sharing between the three types of workers: Management, Production Worker, and Specialist) 1.4  NMT  1.7 (Medium responsibility sharing) 1  NMT  1.4 (Low responsibility sharing) NTM  1 (No responsibility sharing)

Appendix C Logistic Regression Results for the Binary Outcome “Evaluated/Not Evaluated” Table 3C.1

The case of individual production Standard error

Wald χ2

Pr > χ2

–2.0329

0.3485

34.0305

χ2

1.0360 0.6593 1.7797 –0.0104 0.1636 0.0894 0.9187 0.6605 0.3523 0.6672

1.4831 0.2753 0.3673 0.2843 0.2718 0.3122 0.4546 0.2988 0.3630 0.3829

0.4880 5.7342 23.4742 0.0013 0.3625 0.0821 4.0840 4.8870 0.9420 3.0367

0.4848 0.0166 χ2

0.1398 –0.9496

0.1958 0.5843

0.5101 2.6415

0.4751 0.1041

0.1439 –0.1158

0.2051 0.3478

0.4921 0.1108

0.4830 0.7392

–0.1922 –0.0789

0.2010 0.3770

0.9143 0.0438

0.3390 0.8343

–0.2005 0.3645

0.1941 0.3384

1.0673 1.1598

0.3016 0.2815

–0.2209 0.0596 –0.2073

0.1608 0.1448 0.1627

1.8877 0.1694 1.6232

0.1695 0.6806 0.2027

–0.0469 0.1619 0.3439

0.1864 0.2226 0.2386

0.0635 0.5292 2.0764

0.8011 0.4669 0.1496

Estimate

Reference is “men.” Reference is “50 and more.” c Reference is “11 and more.” d Reference is “with no degree except CEP or BEPC.” e Reference is “unskilled blue collar.” f Reference is “99 and less.” g Reference is “food industries.” h Reference is “intensity 1.” See section A of appendix B for the construction of this variable. i The variable is equal to 1 when the firms states that such an obstacle has been either “quite important,” “important,” or “very important,” and 0 when she states that it has been “unimportant.” j Response is either “yes” or “no.” k Reference is “0.” l Reference is “less than 10%.” m Response is either “yes” or “no.” n Reference is “≤ 1.” See section B of appendix B for the construction of the variable. b

Table 3C.2

The case of team production Standard error

Wald χ2

Pr > χ2

–1.7432

0.3745

21.6633

χ2

–0.0540 –0.2445 –0.0976 0.0598

0.1773 0.2312 0.1999 0.1538

0.0927 1.1179 0.2382 0.1513

0.7608 0.2904 0.6255 0.6973

0.0411 0.1569 –0.1195

0.1639 0.1589 0.1930

0.0629 0.9750 0.3833

0.8020 0.3234 0.5358

0.4089 0.1545 –0.0932

0.1604 0.1389 0.1582

6.4964 1.2379 0.3470

0.0108 0.2659 0.5558

0.4285 0.1763 0.1045 –0.3277 0.1577

0.1631 0.1351 0.1380 0.1778 0.1781

6.8979 1.7015 0.5728 3.3980 0.7846

0.0086 0.1921 0.4492 0.0653 0.3757

–0.0279 –0.0204 –0.3818

0.2610 0.1629 0.2959

0.0114 0.0156 1.6652

0.9150 0.9005 0.1969

0.0247 0.0651

0.1780 0.2613

0.0193 0.0620

0.8895 0.8033

0.4672 0.8599

0.1863 0.3590

6.2869 5.7362

0.0122 0.0166

–0.0944 –0.0239

0.1838 0.4492

0.2640 0.0028

0.6074 0.9575

–0.3616 –0.2979

0.1983 0.3314

3.3244 0.8082

0.0683 0.3687

0.1016 0.2644

0.2065 0.3815

0.2419 0.6229 0.4801 0.4884 (continued)

Estimate

130

Marc-Arthur Diaye, Nathalie Greenan, and Michal W. Urdanivia

Table 3C.2

(continued)

Parameter Share of other workers participating in project teams l 10% to less than 50% 50% and more Who is/was authorized in 1997 to participate in project teams? m Management Production worker Specialist Average number of tasks that each type of worker is responsible forn 1.1–1.4 1.5–1.7 1.8 and more

Estimate

Standard error

Wald χ2

Pr > χ2

0.1084 0.0786

0.1851 0.3393

0.3428 0.0537

0.5582 0.8168

–0.3374 –0.1134 0.1610

0.1701 0.1475 0.1591

3.9368 0.5912 1.0239

0.0472 0.4420 0.3116

0.0988 0.1538 –0.1319

0.1917 0.2221 0.2414

0.2656 0.4796 0.2984

0.6063 0.4886 0.5849

a

Reference is “men.” Reference is “50 and more.” c Reference is “11 and more.” d Reference is “with no degree except CEP or BEPC.” e Reference is “unskilled blue collar.” f Reference is “99 and less.” g Reference is “food industries.” h Reference is “intensity 1.” See section A of appendix B for the construction of this variable. i The variable is equal to 1 when the firms states that such an obstacle has been either “quite important,” “important,” or “very important,” and 0 when he or she states that it has been “unimportant.” j Response is either “yes” or “no.” k Reference is “0.” l Reference is “less than 10%.” m Response is either “yes” or “no.” n Reference is “ ≤ 1.” See section B of appendix B for the construction of the variable. b

References Abowd J. M., R. H. Creecy, and F. Kramarz. 2002. Computing person and firm effects using linked longitudinal employer-employee data. LEHD Technical Paper no. 2002–06. Alchian A., and H. Demsetz. 1972. Production, information costs, and economic organization. American Economic Review 62:777–95. Crifo P., M.-A. Diaye, and N. Greenan. 2004. Pourquoi les entreprises évaluent-elles individuellement leurs salariés? (Why do firms evaluate individually their employees?). Economie et Prévision 164–165 (3–4): 27–55. Diaye M.-A., N. Greenan, and M. W. Urdanivia. 2007. Subjective evaluation of performance through individual evaluation interview: Theory and empirical evidence. NBER Working Paper no. 12979. Cambridge, MA: National Bureau of Economic Research.

Subjective Evaluation of a Performance and Evaluation Review

131

Greenan N., and S. Hamon-Cholet. 2001. Un dispositif d’enquêtes couplées employeurs/employés sur les changements organisationnels et l’informatisation (A linked employer-employee survey device on organizational changes and ICT uses). Center for Labor Studies. Mimeograph. Heckman J. J., H. Ichimura, and P. Todd. 1997. Matching as an econometric evaluation estimator: Evidence from evaluating a job training programme. Review of Economics Studies 64:605–54. ———. 1998. Matching as an econometric evaluation estimator. Review of Economics Studies 65:261–94. Rosenbaum P. R., and D. B. Rubin. 1983. The central role of the propensity score in observational studies for causal effects. Biometrica 70:41–55.

4 Do Initial Conditions Persist between Firms? An Analysis of Firm-Entry Cohort Effects and Job Losers Using Matched Employer-Employee Data Till von Wachter and Stefan Bender

4.1 Introduction Economists have long been interested in how persistent the effects of short-term unexpected shocks in the labor market are on workers’ careers (e.g., Okun 1973). Using newly available longitudinal data, an increasing number of papers suggest that the starting conditions in the first year of a worker’s job or labor market entry can indeed have long-term effects on earnings and career development (e.g., Oreopoulos, von Wachter, Heisz 2006; Oyer 2006; Kahn 2006). For example, Oreopoulos, von Wachter, and Heisz (2006) find that the effect of graduating college in a recession fades after ten years for the typical worker and has permanent negative effects for less-able graduates. While clearly a concern for policymakers and the public, such lasting effects of entry conditions are also difficult to explain in the context of standard models of wage setting and career development. In particular, they raise the question of whether wages persistently deviate from workers’ skills because of market frictions or wage contracts. This question has received particular attention in the context of cohort Till von Wachter is an assistant professor of economics at Columbia University, and a faculty research fellow of the National Bureau of Economic Research. Stefan Bender is a senior researcher at the Institute for Employment Research. We would like to thank David Card, Bob Gibbons, Larry Katz, and conference participants at the Comparative Analysis of Enterprise Data conference (CAED) 2006 in Chicago and the Conference on Analysis of Firms and Employees (CAFE) 2006 in Nuremberg for helpful suggestions. Ana Rute-Cardoso provided helpful comments. This chapter was written as part of the research project “Discrepancies between Market and Firm Wages: An Analysis of Earnings and Worker Mobility” within the German Research Foundation’s (DFG) research program “Flexibility in Heterogenous Labor Markets” (SSP 1169). This material is based upon work supported by the National Science Foundation under grant number 0453017. All errors are our own.

135

136

Till von Wachter and Stefan Bender

effects within firms. A small but influential number of papers have argued that similar workers entering firms in different years receive permanently different wage profiles (Baker, Gibbs, and Holmstrom 1994; Beaudry and DiNardo 1991). Several approaches have been proposed to rationalize such persistent shifts in firms’ wage structures. The first maintains that the degree of rent sharing between workers and firms varies with outside market conditions at the time of entry (Beaudry and DiNardo 1991). The second maintains that cohort effects arise from variation in the quality of jobs and career opportunities available within the firm (Okun 1973). If different jobs provide different general experience or training provided by the firm, cohort effects can also arise from permanent changes in workers’ skills (Gibbons and Waldman 2004). Although these explanations have very different underlying views of wage determination, they have similar predictions for the degree of persistence of entry-level conditions. Thus, it is difficult to distinguish among them based on cohort effects in earnings alone. However, these explanations have alternative implications for the persistence of entry-level conditions as workers switch employers. While effects due to rent sharing or job quality should fade for those workers losing their jobs, changes in skills should affect workers’ wages even at new employers. Despite offering clear predictions, these hypotheses have not been tested, in part because data used in existing work had little information on workers’ job mobility and their employers. More generally, because existing studies focused on single firms (Baker, Gibbs, and Holmstrom 1994) or particular time periods (Beaudry and DiNardo 1991), at present little is known about whether firm-entry cohort effects are a pervasive phenomenon in the wider labor market. Given the degree of heterogeneity in other aspects of firms’ wage structures (Abowd and Kramarz 1999) and given the amount of heterogeneity in firm growth rates (Davis and Haltiwanger 1992), it is conceivable that firm-entry cohort effects are a widespread phenomenon that affects firms to different degrees. However, until now little information is available on how pervasive such cohort effects are. In this chapter, we provide three contributions to the present literature. First, we use data on the complete career histories of all workers in a large German manufacturing sector to describe the prevalence and heterogeneity of firm-entry cohort effects for a large sample of firms over more than twenty years. To ensure the cohort differences in wages we find are not due to selective entry of workers into firms, the nature of our data allows us to control for observable firm and worker characteristics as well as worker fixed effects. In addition, the long time horizon allows us to examine whether entry conditions fade within firms and whether firms’ wages tend to converge to a common market wage over time. Second, we exploit the predictions of the alternative models for the im-

Do Initial Conditions Persist between Firms?

137

pact of job loss on wages to learn more about the sources of firm-entry cohort effects. To do so, we complement the descriptive analysis with a study of the effects of job displacement on wage changes for workers with high, medium, or low starting wages at the lost job. Thereby, we are particularly interested in whether wage premiums fade upon job loss and whether workers recover some of their past advantages with time since job loss. Third, we analyze the effect of past wage premiums on the level of wages after job loss. Because controlling for observable characteristics past wages are partly a function of unobserved ability, we would expect a positive correlation. However, if the ability of job losers is not observed perfectly by the market, temporary wage premiums may also serve as a temporary signal that fades over time. If, on the other hand, wage premiums are driven by permanent skill differences, we would expect their effect to be stable or increasing. We find that in the manufacturing sector we study, firm-entry cohort effects are a significant phenomenon. Similar firms pay different wages to similar workers starting their jobs at different points in time. However, we also find that this is not simply a homogeneous market-wide phenomenon—there is considerable heterogeneity between firms and between cohorts in the incidence and strength of cohort effects. A further key result is that in our sample, entry-level differences in wages fade within firms, and there appears convergence to a market wage, but reversion is very slow. Thus, wage differences between cohorts of similar workers are highly persistent but not permanent. We also find that workers with high starting wages have higher and persistent wage losses at job loss; workers with relatively low starting wages, on the other hand, seem to gain from losing their job. Thus, part of initial wage differences appears to be temporary firm-specific rents. Moreover, there appears to be mean reversion at job loss. However, markets do not seem to be able to fully tell apart ability from rents in the short run, and past wage advantages carry a premium for the level of wage after job loss that fades over time. These results suggest that firm-entry cohort effects at least in part consist of time varying differences in rent sharing or job quality. Clearly, part of the effects we find may also arise due to the presence of other individual specific rents, for example, from job search. Future research based on a larger sample of firms and workers able to explicitly analyze the persistence of cohort effects at job loss will help to shed light on this question. The results also suggest that characteristics of the previous job, such as job tenure or past wages, are not just a fixed measure of worker quality, as suggested in the prior literature (e.g., Kletzer 1989) but also appear to influence temporary wage components. Among others, this could arise if previous job characteristics affect workers’ reservation wages. The effect of these initial conditions fades, consistent with the notion of continued on-the-job

138

Till von Wachter and Stefan Bender

search. In addition, previous job characteristics may function as temporary signals of workers’ skills.1 The outline of the chapter is as follows. First, we give a brief overview of the conceptual background, the empirical approach, and the data we use. Second, we describe the prevalence of cohort effects in a sample of large and stable manufacturing firms. Third, we analyze the effect of past starting wages on the extent of wage changes at job displacement. Fourth, we study the effect of the starting wage on the lost job on the level of ensuing wages. The last section concludes and offers suggestions for future research. 4.2 Conceptual Approach There are two basic explanations for the persistence of differences in starting wages of workers entering the same firm at different moments in time. The first view suggests that wages contain firm-specific components that can differ across entry cohorts but that are lost as workers move between firms. This may arise due to differences in the degree of rent sharing among workers and firms, for example, due to the degree of pressure in the outside labor market. Or it may arise to the presence of long-term implicit insurance contracts (Beaudry and DiNardo 1991). Alternatively, this may be due to variation in the quality of jobs offered within firms over time (Okun 1973). For example, in periods of high growth, firms may offer more jobs that pay more, either because of higher productivity or due to higher incentive wages. Persistent differences may also arise if some jobs provide higher accumulation of firm-specific skills. These alternative sources of wage differentials have the similar implication that the wage advantages they may imply for certain cohorts are lost if workers leave the firm. Because voluntary movers may not leave their job if compensated for giving up of these wage premiums, the loss is likely to be visible only for workers who move their job involuntarily. Thus, we would expect wage losses for those job losers to be largest that had the highest wage premiums. For these displaced workers, we would expect to see mean reversion; that is, those workers with below-average cohort wages experience wage gains relative to those workers with above-average cohort wages as absent any skill differentials, both groups draw again wages from the same market wage distribution. Because the workers with below-average cohort wages could have obtained higher wages on the outside market, some mobility friction must prevent them from moving jobs. Because cohort effects are likely to be more typically in large firms with longer job attachment, this is likely to arise due to the presence of average wage premiums large firms pay (Oi and 1. However, in that case the effect of the initial signal should not fade over time (Farber and Gibbons 1996; Altonji and Pierret 2001).

Do Initial Conditions Persist between Firms?

139

Idson 1999). Nevertheless, we would expect that on average workers with below-average cohort wages are more like to switch employers. Similarly, firms may face an incentive to fire workers with above-average wages if these are due to a higher amount of rents. The second broad view suggests firm-entry cohort effects arise from changes in workers’ general skill level. This may occur if in some periods firms offer a larger amount of jobs with a high degree of experience accumulation or general training (Gibbons and Waldman 2004). In this case, differential entry-level conditions reflect actual differences in workers’ skill levels and can arise even in an environment where each worker is paid his marginal product. This is in contrast with the first set of explanations, that each suggested that workers with similar skills would be paid different wages, either because of rent sharing or differences in job quality. Clearly, the second view suggests that even workers losing their job involuntarily will maintain their wage advantage on their new job at least in the medium run. While in the years immediately following the job loss some of the advantage may be lost as workers have to find a new job match or as the market may be uncertain about workers’ ability, in the medium run, workers should again obtain a wage that reflects their higher (or lower) marginal product. This stands in contrast to the implications of the first view, in which all cohort-wage differences should be lost at job loss. In particular, even if past wages may serve as a positive signal for ability in the years immediately after job loss, the effect of past cohort conditions should fade with time since job loss—the opposite implication as from the second view. The existing empirical literature does not address the question of persistence of conditions on the past job for workers switching employers. One strand of literature aims at characterizing the presence of firm-entry cohort effects, but pays little attention as to what happens when workers leave firms. In this vein, Baker, Gibbs, and Holmstrom (1994) analyze the role of cohort effects within a single firm. Beaudry and DiNardo (1991) use data from the Current Population Survey (CPS) and Panel Study of Income Dynamics (PSID) to analyze the effect of labor market conditions on workers’ wages as they stay within the firm. Neither paper analyzes the persistence of the wage effect it finds as workers move between firms, mostly due to a lack of data. Another strand of literature examines the extent and determinants of wage changes at job loss in detail, but typically pays less attention to the role of past job characteristics.2 The only important exception is the role of past job tenure. Because there is no market for firm worker-specific skills or match rents, the wage gradient with job tenure can be seen as a form of 2. Past industry, occupation, and firm size are an exception. See, for example, Ruhm (1991). Jacobson, LaLonde, and Sullivan (1993), Gibbons and Katz (1991), or Farber (1997, 2003). For a survey of this literature, see Farber (1999).

140

Till von Wachter and Stefan Bender

rent sharing between workers and firms. A worker losing his job should then lose these firm-specific rents. This is what the literature has found, and the effect appears to be particularly strong for a loss in industry tenure (Neal 1995; Parent 2000). In this context, Kletzer (1989) has found that workers with higher past job tenure have higher wages on the job after job loss. This may signify that workers with high job tenure are also of high ability, that is, positive wage tenure profiles in part reflect ability differences between high- and lowtenured workers.3 A similar argument holds for the effect of the initial wage on the lost job. Even conditional on observable characteristics—such as age and education—past starting wages will be a function of unobserved worker ability and will thus positively correlate with wages on the current job. However, past tenure and earnings may also influence workers’ reservation wages. In this case, high past wages may lead workers to search for jobs more intensely. If this is the case, there is again an initial correlation of past job characteristics and initial earnings after layoffs. Over time, these workers’ wages are again determined by market conditions (workers’ skill levels and the overall wage distribution); thus, the effect of the reservation wage would be expected to fade. In addition, if the market observes workers’ ability only imperfectly, it may use past job tenure or past wages as signals to infer about their productivity (Farber and Gibbons 1996; Altonji and Pierret 2001). In this case, part of the positive effect of past job tenure may be due to an initial signaling effect. However, this effect should not fade over time, even if markets learn about workers’ ability. If, on the other hand, firm-entry cohort effects are due to differential skill accumulation, we should observe the opposite phenomenon. Initially, some of the higher skills embodied in the cohort-effect may be discounted if displaced workers receive a wage based on average skills. Over time, as markets learn about workers’ true ability, we would expect the effect of past wages to remain stable, or least not to decline further. 4.3 Empirical Approach The analysis of the chapter consists of two parts, each based on a different sample of firms. The first, descriptive part of the paper studies the importance of firm-entry cohort effects for a sample of large stable firms in the car manufacturing sector in Germany. The second part analyzes wage changes and wage levels of job losers using the complete available career histories of all workers who ever worked in German car manufacturing. 3. This idea is also exploited in Abraham and Farber (1987), who use completed job tenure as an indicator for the quality of a job match to correct for selection bias in estimates of the return to job tenure.

Do Initial Conditions Persist between Firms?

141

The data is drawn from the German employee registry that records complete career information as well as basic demographics for the universe of German workers covered by social security and their employers from 1975 to 2003 and is further described in the following. The goal of the first part of the paper is to describe the incidence, heterogeneity, and persistence of firm-entry cohort effects within a large but specific sector of the economy. The focus on a single sector allows us to exclude wage differences arising from differential industry trends or business cycles. To study the magnitude and evolution of average cohort wages, we concentrated our analysis on stable establishments with a large enough rate of inflow of new workers in every period. For each of the fifty-five firms that survive our selection criteria further described in the following, we estimate cohort effects following the approach in Baker, Gibbs, and Holmstrom (1994). To do so, we proceed in three steps. First, we collapse our data to the level of firm-tenure entry year cells. Second, we use the cell-level averages to run the following wage regression at the firm level. (1)

lo gw fct  f  gf (ten)  ft  cf  f X fct  ufct

This modeling approach allows for a firm-specific quartic tenure profile [gf (t)], a constant and year effects, as well as for firm-specific effects of average entry cohort characteristics. Third, we regress the estimated firmentry cohort effects (cf ) on a firm-specific trend and treat the residual from that regression as cohort effects for the remainder of our study. As explained in Baker, Gibbs, and Holmstrom (1994) in the presence of year and tenure effects, one cannot identify the linear component of the cohort effect. Because we are mainly interested in examining the presence and significance of cohort effects, the chosen approach suffices for our purposes. In addition to including average observable characteristics at the cohort level, we also ran the model in equation (1) at the individual level and included worker fixed effects. Unlike in the case of Baker, Gibbs, and Holmstrom (1994), who only had access to data on all workers at a single firm, this is possible in our case because we have the entire career information of workers who ever worked at each of our firms. This further alleviates the concern that the cohort effects identified in equation (1) may still be due to selective entry of workers of different skill levels. An important aspect of firm-entry cohort effects is their persistence— do differences in entry-level wages last unfettered forever, as found in the firm analyzed by Baker, Gibbs, and Holmstrom (1994), or does convergence take place? Convergence may be of two kinds. First, high-wage cohorts may converge to the average-wage level within the firm. In this case, the relevant benchmark and speed of convergence is determined by the firm-level average. Second, high-wage cohorts may converge to a marketlevel wage. That is, reversion of high initial starting wages may be faster if they are high relative to the overall market wage.

142

Till von Wachter and Stefan Bender

To examine the extent and speed of reversion of initial wage differences, we modify the preceding model and estimate the following regression for each firm in our sample of large stable firms. (2)

lo gw fct  f  gf (ten)  ft  f X fct  cf 0  hf (ten)cf 1  ufct

Thereby, cf 0 measures the difference in initial starting wages for entry cohort c, and cf 1 measures the firm-specific rate of decay of the initial effect. We experimented with linear, quartic, and unrestricted specifications for the decay function hf (t), and found a linear specification works astonishingly well for the most relevant time horizon of about ten years of job tenure. The second part of the paper studies the effect of starting wages on the effect of job displacements. Once we have identified displacement events and an appropriate estimation methodology, the analysis is relatively straightforward. In particular, we are interested in whether wage losses at job displacement differ by the level of the starting wage at the previous job. Ideally, we would have analyzed the effect of firm-entry cohort effects themselves on the extent of wage loss for workers losing their jobs from our sample of large and stable firms. However, for the sector in question, the sample of such workers was too small for a meaningful analysis. Thus, in the second part of the paper, we analyze the effect of a job displacement on wage changes and post-job loss wage levels for all workers who worked in German car manufacturing at some point between 1975 and 2003. We define a displaced worker to be a worker who had at least three (or five) years of tenure at a given firm and who had at least thirty days of unemployment following the job move. We experimented with alternative definitions based on mass layoffs at the establishment level, but again found that we had too few workers affected by such events in our sample.4 We then study the wage change of displaced workers relative to the wage held prior to job loss for up to fifteen years after the job change. Specifically, the basic model we estimate at the individual level is 15

(3)

log wit  i  ∑ Dkit k  g(expit)  t  uit, k–3

where the dummies Dik indicate whether a year is k periods before or after a job loss, and y stands for calendar year. This estimates the effect of wage changes at job loss controlling for a quartic polynomial in potential labor market experience, unrestricted year effects, and worker fixed effects. This model essentially extends Farber’s estimates based on the Displaced 4. In a separate work using the German Socio-Economic Panel, Görlitz and von Wachter (2006) find that while imposing unemployment does tend to raise the estimated impact of job losses relative to self-reported layoff status, the difference is reduced significantly when worker fixed effects are included.

Do Initial Conditions Persist between Firms?

143

Worker Survey (DWS) Supplement to the CPS into an analysis covering several periods after the job loss. In particular, this approach does not keep a control group of workers who did not lose their job, and thus differs from the estimation method implemented by Bender et al. (2002) for Germany, based on Jacobson, LaLonde, and Sullivan (1993). Instead, the year effects in this sample are identified from the baseline period of workers later experiencing displacement.5 The main estimates we are interested in are estimates of the earnings loss by groups of workers with low, medium, and high starting wages relative to their average wage. Thus, we reestimate the model in equation (3) interacting the time effect as well as the displacement-time effects with dummies for whether a worker’s starting wage at the lost job was in the bottom, middle, or top of the wage distribution (we choose the interquartile range as cutoff points). This results in the following model for estimation: 15

(4)

15

15

k–3

k–3

log wit  i  ∑ D k,Low Low  ∑ D k,Med Med  ∑ D k,High High it k it k it k k–3



Low t



Medium t



High t

 g(expit)  uit

The estimates of this model show the wage changes by groups of workers with different starting wages relative to their own group-specific wage at the time of job loss. In future work, we plan to include stayers—workers who did not lose their job—in the model as a control group to replicate the classic event study design introduced by Jacobson, LaLonde, and Sullivan (1993). We will also analyze wage losses by other worker characteristics such as education, age, or past job tenure. Similarly, we can exploit further prediction regarding the effect of past job characteristics on the wage changes of voluntary movers. The last set of models we estimate focus on the level of log wages after job loss. To do so, we begin by implementing the models estimated by Kletzer (1989), who concentrates on the effect of past job tenure. We first augment Kletzer’s model with the effect of past starting wages. Then we extend her approach and interact past job tenure and past starting wages with time since job loss. Thus, we are interested in the coefficients on the interactions with time since job loss in the following model: (5)

log wit    0log wLostJob  1 log wi0LostJob  Xi i0  g(expit)  t  εit,

5. To identify the worker fixed effects, we have to exclude one pre-period dummy. To identify the year effects, we have to exclude one additional dummy. Thus, we keep observations on workers up to five years prior to displacement and include dummies for up to three years prior to displacement.

144

Till von Wachter and Stefan Bender

where stands for the years since job loss. This model is only estimated based on observations after a job loss. The important extension of Kletzer’s model is made possible by the availability of longer time series in our data and allows us to study to what extent the immediate effect of past job and worker characteristics on wages post job loss fades over time. Alternatively, we will be able to see whether past wages are driven by components of actual or predicted worker skill whose effect stays stable. 4.4 Administrative Longitudinal Matched Data The data used in this chapter are drawn from the German employment register containing information on all employees covered by social security, representing around 80 percent of the German workforce.6 The employment register takes stock of existing employees at each establishment twice a year. Because the notification procedure for social security also requires employers to record any permanent or temporary change of employment relationships, the employment register contains detailed histories for each worker’s time in covered employment. The main information contained in the register for administrative purposes (and, therefore, the most reliable) are gross daily wages subject to social security contributions and the exact periods during which the employee worked in the social security system. In addition, the data contain basic demographic information as well as information on occupation, industry, job status, and education.7 Most important for the present purpose, the data also contain unique establishment identifiers. These were used to create a separate data set of establishment characteristics that were aggregated up from the employment register and merged back onto the individual-level data. Characteristics include, among others, establishment size, employment growth, and average wages. The relevant entity throughout the empirical analysis is the establishment. Despite the inaccuracy it entails in some cases, we will keep using the terms establishment and firm interchangeably for the rest of the analysis.8 The sample used for this chapter consists of information on the universe of workers and establishments from the West German car-manufacturing 6. An overview of the data is given in Bender, Haas, and Klose (2000); a more detailed description can be found in Bender et al. (1996). For further information and citations as well as accessibility, see http://www.research-data-center.de. Coverage includes full- and part-time employees of private enterprises, apprentices, and other trainees, as well as temporarily suspended employment relationships. The self-employed, civil servants, and students are excluded. 7. The entity reporting is the establishment for which an employee works and can thus change over time. This can lead to mistakes in the coding of some demographic variables (e.g., nationality or marital status) and in particular education (which tends to reflect required rather than actual qualification). 8. Unfortunately, it is currently not possible to link establishments that belong to a common parent firm.

Do Initial Conditions Persist between Firms?

145

sector. In a first step, we selected all employees who worked at least one day between 1975 and 2003 in an establishment of this sector (a total of 162,332 establishments). To ensure that the sample is consistent in time, we chose only those notifications where the employees worked part or full time. We dropped apprentices from the main analysis to avoid confounding job changes at end of apprenticeship with regular job displacement and to be consistent with the concept of firm-entry cohort effect typically analyzed in the literature. We also dropped workers with missing education and who are younger than twenty-one and older than sixty-four. Using this sample, we aggregated up the individual-level information into a cell-level data set at the establishment, year, and entry cohort level that contains the size of each entering cohort in each year at the firm, as well as average earnings and basic average demographic characteristics (such as average age, average education, or fraction female). To obtain a meaningful basis for the descriptive analysis of firm-entry cohort effects, from this cell-level data set we extracted a subset of firms that had a sufficiently large inflow of workers each year for an extended period of time. In particular, we required firms to have at least ten entering cohorts with at least ten employees, at least 100 employees over ten years, and at least twenty-one entering cohorts. This leaves us with a total number of fifty-five firms. This restriction ensures both a reasonable sample of firms as well as a meaningful base for calculation of a large number of firm-entry cohort effects. We have experimented with the cutoff points, without a noticeable difference in results. In addition, to ensure we observe each cohort for an extended amount of time, we only consider cohorts entering before 1997. For the displacement analysis, we selected from our sample of car manufacturing all workers with at least three years of tenure who changed employers and who spend at least thirty days in unemployment after moving. For this sample, we only kept observations that were at least five years before and at most fifteen years after the job loss. Characteristics of various samples of displaced workers are shown in table 4.1. 4.5 Empirical Results 4.5.1 Firm-Entry Cohort Effects in German Car Manufacturing To illustrate our main descriptive results, we begin by showing the pattern of firm-entry cohort effects for a single large and stable establishment in the car manufacturing sector.9 Figure 4.1 shows the development of average log real daily wages for biannual entry cohorts ranging from 1976 to 1996. One can clearly see a rising trend and significant fluctuations in en9. For data protection reasons, we have added random constant with zero mean to the individual wage levels.

146 Table 4.1

Till von Wachter and Stefan Bender Sample characteristics of stable firms and displaced workers in WestGerman car manufacturing, 1975–2003 A. Basic characteristics of 55 stable and large firms in car manufacturinga

Number of cohorts Employment size Size of entry cohort Average cohort age Average cohort fraction female Average cohort years of education Average cohort starting wage Average cohort log real daily wage

Average

SD

Median

19.6 6376.7 482.3 39.8 0.11 10.50 4.33 4.49

4.4 9560.8 1678.8 6.3 0.05 0.63 0.10 0.17

22.0 2161.0 88.0 40.5 0.10 10.25 4.31 4.49

B. Average characteristics of various samples of displaced workersb

Years of Job Tenure Prior to Job Loss

Three

Five

Three, From 55 Large Stable Firms

Fraction female Fraction non-German Years of education Average age Average potential experience Average tenure on lost job Fraction part-time on lost job Fraction low-skill blue collar on lost job Fraction high-skill blue collar on lost job Fraction low-skill white collar on lost job Average log real daily starting wage Average log real daily wage

0.14 0.17 10.45 35.30 18.85 5.57 0.03 0.37 0.43 0.18 4.25 4.13

0.14 0.18 10.39 37.23 20.84 7.70 0.03 0.38 0.42 0.18 4.26 4.15

0.14 0.21 10.21 34.40 18.19 3.31 0.08 0.50 0.49 0.29 4.25 4.28

a

Statistics based on firm-year-cohort observations or averages. Average cohort characteristics are weighted by cohort size. b Sample only includes observations for workers who moved jobs followed by a spell of thirty days of unemployment or more at least once. Averages are taken over workers and workeryears ranging from five years before to fifteen years after job loss.

try wages over time. More important, the difference in entry wages clearly leads to persistent average-wage differences across cohorts. However, the figure also clearly shows a pattern of reversion. Differences in initial wages appear to fade over time. These patterns are documented explicitly in figure 4.2, which shows the annual entry-cohort effects obtained by estimating equation (1) and detrending the resulting cohort effects. One can clearly see permanent differences in average wages of different firm-entry cohorts. Controlling for observable characteristics reduces the cohort effects only somewhat. This suggests that when the firm pays higher wages, it attracts more able workers. However, if we instead control for worker fixed effects, the cohort wage

Fig. 4.1

Average wages by biannual entry cohorts for a single firm

Note: For data protection reasons, we have added random constant with zero mean to the individual wage levels.

Fig. 4.2

Firm-entry cohort effects and starting wages for a single firm

148

Till von Wachter and Stefan Bender

differences seem to rise, leaving us with no clear conclusion regarding selective entry between cohorts. In either case, we find there are robust differences in average cohort wages over time in this large manufacturing firm, as suggested by Baker, Gibbs, and Holmstrom (1994) for a large financial service firm. However, contrary to the finding in Baker, Gibbs, and Holmstrom (1994), the pattern in the figure also shows that average cohort differences in wages are smaller than differences in average starting wages between firm-entry cohorts. Figure 4.3 shows the time pattern of reversion of initial wage differences explicitly for different specifications of the decay function. Unlike Baker, Gibbs, and Holmstrom (1994), we find a significant albeit slow decay of initial wage differences that lasts up to twenty years. Perhaps not surprisingly, we find a concave tenure wage profile (the profile in Baker, Gibbs, and Holmstrom’s [1994] firm was linear). The pattern of decay we find is approximately linear. The key question then is to what extent the result of statistically and numerically significant firm-entry cohort effects hold for a wider sample of firms as well. The answer to this question is affirmative. We ran the model in equation (1) separately for each firm in our sample of fifty-five large and stable car manufacturing firms and detrended each set of cohort effects as described in section 4.3. The distribution of estimated cohort effects for all firms is shown in figures 4.4 and 4.6 and table 4.2. Figures 4.5 and 4.7 show the distribution and the reversion of initial wage differences.

Fig. 4.3

Decline in effect of entry wages with tenure at firm

Fig. 4.4

Distribution of firm-entry cohort effects in different years

Fig. 4.5

Distribution of average firm-entry wages: Different years

Table 4.2

Firm-entry cohort effects and average starting wages in German car manufacturing, 1975–2003 Without worker characteristics (year-group) 1980s

Controlling for worker characteristics (year-group)

Percentile

1970s

1990s

1970s

1980s

1990s

10 25 50 75 90

A. Distribution of firm-entry cohort effects by decade –0.041 –0.062 –0.059 –0.041 –0.062 –0.014 –0.030 –0.027 –0.014 –0.030 0.003 –0.004 –0.003 0.003 –0.004 0.020 0.022 0.029 0.020 0.022 0.052 0.052 0.067 0.052 0.052

–0.059 –0.027 –0.003 0.029 0.067

10 25 50 75 90

B. Distribution of average starting wages of firm-entry cohorts –0.046 –0.115 –0.112 –0.043 –0.094 –0.019 –0.058 –0.060 –0.017 –0.048 0.009 –0.014 –0.009 0.005 –0.011 0.035 0.021 0.035 0.028 0.018 0.079 0.064 0.092 0.062 0.051

–0.098 –0.047 –0.006 0.029 0.068

Notes: Distribution of average cohort wages by year-group. All models estimating cohort effects shown in panel A also include a firm-specific quartic tenure profile, firm-specific year effects, and a firm effects. The resulting firm-entry cohort effects are detrended. Average starting wages are net of year effects and firm effects. The observable characteristics in the right-hand panels are fraction female, fraction non-German, fraction without degree, fraction with apprentice degree, fraction with college degree, fraction low-skilled or high-skilled blue collar, and fraction low-skilled white collar. All models are weighted by the cohort size.

Fig. 4.6

Percentiles of fraction of initial wage effect decayed

Do Initial Conditions Persist between Firms?

Fig. 4.7

151

Decay of initial effect by percentile of average starting wage

Overall, we obtain five core results: 1. There are significant cohort effects for each firm in the industry we study that are robust to controls for worker and firm characteristics. Similar workers entering firms at different times earn different wages. 2. There is considerable heterogeneity of cohort effects between firms. The entry cohort effects cannot be simply driven by overall labor market conditions in the industry. 3. Heterogeneity in cohort effects (both within and between firms) is increasing over time. The spreading of the German wage distribution occurs in part through cohort effects. 4. Cohort wage differences are largest for entry-level wages and converge over time within firms. However, convergence within firms is slow, such that persistent differences in average wages remain. 5. Reversion of wages is faster the farther average cohort wages are from the overall market. Outliers tend to convergence between firms as well. The distribution of cohort effects with and without worker characteristics is shown in table 4.2 for the full sample and each of the three decades of our sample. The distribution of F-statistics or p-values is omitted because all cohort effects are significant at the 1 percent confidence level. The table also shows the distribution of average entry-level wages with and without worker controls. The results suggest that there are important and significant differences in average wages of firm-entry cohorts that are robust to controls for average worker characteristics.

152

Till von Wachter and Stefan Bender

The typical detrended cohort effect lies within plus and minus 5 percent of average firm wages. Taken at face value, they suggest that some cohorts in some firms carry premiums or discounts on the order of 5 percent, which corresponds to the wage effect of about one year of labor market experience or a year of education in Germany. Given we cannot identify the linear component of cohort effects, care should be taken with interpreting the specific magnitudes. The average differences in cohort effects mask even bigger differences in average starting wages between cohorts. Comparing figures 4.4 and 4.5, one can see that the distribution of deviations of cohorts’ starting wages from firm-specific averages has fatter tails. Again, although most of the differences are limited in magnitude, some cohorts experience large differences in average wages. The distribution of cohort effect arises from differences between cohorts within firms. However, a large part of the variation arises from variation between firms for any given cohort. This is apparent from the fact that it holds within decades and can be shown to hold within single years as well. In fact, the annual distribution of cohort effects is similar to the decadewide distribution, suggesting that an important part of the variation is coming from between firms. Thus, firm-entry cohort effects cannot be simply explained by business-cycle pressures affecting the entire industry. It may be that within the industry, firms producing different products (say trucks or passenger cars) or goods of different qualities face differential demand conditions. In addition, there may be truly firm-specific differences in the evolution of productivity, employment, and output that affect the fortunes of workers entering firms at different points in time. That similar firms within sectors can experience vastly heterogeneous patterns of employment growth has been suggested in the literature before (e.g., Davis and Haltiwanger 1992). Our findings suggest that such differences can lead to differences in entry-wage levels and average wages between entry cohorts and between firms. Interestingly, the numbers in the tables and figures suggest that the distribution of entry wage differences and cohort effects has been widening over time. The increasing spread is consistent with a widening in the German wage distribution in the 1990s after a period of relative stability. Our results suggest that part of the recent widening is due to an increasing spread in entry-wage differences. However, our results also suggest that this pattern had already started in the 1980s, something typically not found in analyses of the overall wage distribution. Figure 4.6 shows the distribution of the fraction of the initial difference in average starting wages decayed at each tenure year. The figure suggests first, that the median rate of decay is very slow, leading to a half-life at about eight to nine years. Second, the figure shows that the speed of decay

Do Initial Conditions Persist between Firms?

153

varies widely between firms. For the bottom decile, the entry-wage difference fades within three to four years; for the top 10 percent, the effect actually increases over time. Convergence does not only occur within firms. Figure 4.7 shows that cohorts that have high average wages relative to the overall market have faster speed of convergence. Thus, convergence also occurs between firms toward the average wage in the market. Overall, these results suggest that firm-entry cohort effects are a significant phenomenon in a broad sample of large and stable manufacturing firms even when controlling for worker characteristics. There is substantial heterogeneity in cohort effects between firms. Convergence within and between firms occurs but is slow. These preliminary estimates suggest that firms’ wage structures have a component that systematically varies over time and differs between firms. Our documentation of this dynamic component complements and extends existing characterizations of static differences in average wages, tenure-wage profiles, and the variance of wages (e.g., Abowd and Kramarz 1999; Abowd, Corbel, and Kramarz 1999; Margolis 1995). These results also underline the importance of efforts to understand the empirical sources of firm-entry cohort effects and their theoretical underpinnings. The descriptive results in the previous section allow no clear interpretation with respect to the source of cohort effects. On the one hand, the fact that initial wage differences fade suggests that they must have at least in part been driven by temporary differences in cohort-specific rents or job quality. However, the high degree of persistence does not exclude that some of the effect is driven by lasting differences in workers’ skill levels. This underscores the need of an explicit test of potential explanations that goes beyond purely descriptive study of wage differences themselves. 4.5.2 Job Losses and Differences in Starting Wages As discussed at the outset, if differences in cohort wages arise due to differences in temporary rent or job quality, they should fade if workers lose their job. Alternatively, if cohorts obtain a different degree of training or experience, they should carry their higher skills over to their new job. As discussed at the outset, we study this question by comparing the wage losses of job losers with high or low starting wages at the lost job. The analysis of losses in cohort effects per se is left for future work with a larger sample of firms and workers. As a first step, table 4.3 shows the overall effects of job displacements on wage changes. The time pattern before and after job loss is shown with standard error bands in figure 4.8. The results indicate significant and large wage losses of about 10 percent in the first year that fade in about six to seven years. These results are quite similar to estimates of the effect of job loss in the United States based on the DWS (e.g., Farber 1997, 2003), and similar to estimates in Couch (2001) using a similar methodology and the

Table 4.3

Year to job loss –3 –2 –1 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 Constant No. of observations R2

Wage losses at job loss fifteen years post-job loss, different samples and by starting wage at lost job

Three years of pre-job loss tenure

Five Years of pre-job loss tenure

Exits from large and stable firms

0.0043 (0.0037) 0.0053 (0.0058) 0.0031 (0.0080) –0.0331 (0.0103) –0.1048 (0.0126) –0.0999 (0.0148) –0.0938 (0.0171) –0.0791 (0.0193) –0.0689 (0.0216) –0.0596 (0.0239) –0.0515 (0.0261) –0.0441 (0.0284) –0.0346 (0.0307) –0.0212 (0.0330) –0.0019 (0.0353) 0.0164 (0.0375) 0.0265 (0.0398) 0.0360 (0.0421) 0.0449 (0.0444)

–0.0002 (0.0039) –0.0087 (0.0061) –0.0221 (0.0083) –0.0624 (0.0106) –0.1745 (0.0131) –0.1869 (0.0153) –0.1919 (0.0176) –0.1873 (0.0199) –0.1902 (0.0223) –0.1890 (0.0246) –0.1909 (0.0270) –0.1909 (0.0293) –0.1868 (0.0316) –0.1843 (0.0340) –0.1670 (0.0363) –0.1562 (0.0387) –0.1523 (0.0410) –0.1504 (0.0434) –0.1495 (0.0458)

0.0049 (0.0157) 0.0049 (0.0244) 0.0011 (0.0336) –0.0313 (0.0430) –0.3397 (0.0533) –0.3419 (0.0620) –0.3468 (0.0713) –0.3244 (0.0807) –0.3256 (0.0901) –0.3048 (0.0995) –0.2867 (0.1089) –0.2910 (0.1183) –0.2650 (0.1278) –0.2602 (0.1373) –0.2580 (0.1467) –0.2317 (0.1563) –0.2274 (0.1657) –0.1953 (0.1753) –0.2034 (0.1847)

3.558 (0.0185) 501,103 0.61

3.595 (0.0239) 284,297 0.62

3.670 (0.0980) 25,059 0.63

Wage loss by interquartile range of starting wage of lost job, Three years pre-tenure Low

Medium

High

–0.0161 (0.0048) –0.0294 (0.0063) –0.0118 (0.0082) –0.0378 (0.0103) 0.0794 (0.0126) 0.0998 (0.0146) 0.1193 (0.0167) 0.1352 (0.0188) 0.1517 (0.0210) 0.1571 (0.0231) 0.1644 (0.0253) 0.1722 (0.0275) 0.1728 (0.0297) 0.1877 (0.0318) 0.2002 (0.0340) 0.2087 (0.0362) 0.2163 (0.0384) 0.2209 (0.0406) 0.2282 (0.0428)

0.0071 (0.0039) 0.0114 (0.0058) 0.0039 (0.0078) –0.0372 (0.0100) –0.1513 (0.0123) –0.1483 (0.0144) –0.1436 (0.0165) –0.1301 (0.0186) –0.1218 (0.0208) –0.1132 (0.0230) –0.1066 (0.0252) –0.1005 (0.0273) –0.0929 (0.0295) –0.0855 (0.0317) –0.0748 (0.0339) –0.0607 (0.0361) –0.0566 (0.0383) –0.0525 (0.0405) –0.0482 (0.0427)

0.0188 (0.0046) 0.0366 (0.0062) 0.0221 (0.0082) –0.0192 (0.0102) –0.2139 (0.0128) –0.2205 (0.0146) –0.2251 (0.0167) –0.2145 (0.0188) –0.2112 (0.0210) –0.2028 (0.0231) –0.1949 (0.0253) –0.1929 (0.0275) –0.1820 (0.0297) –0.1747 (0.0319) –0.1530 (0.0341) –0.1356 (0.0363) –0.1266 (0.0385) –0.1203 (0.0407) –0.1192 (0.0429)

3.582 (0.0178) 501,103 0.64

Notes: The entries in the tables are coefficient estimates of regressions of log daily real wages on displacement indicators interacted with dummies for years before and after job displacement. The omitted category are years four and five before job loss. All models also include individual fixed effects, year fixed effects, and a fourth order polynomial in potential labor market experience. The sample excludes apprentices. Standard errors clustered at the individual level are in parentheses.

Do Initial Conditions Persist between Firms?

Fig. 4.8

155

Wage loss for workers losing jobs in car manufacturing

German Socio Economic Panel (GSOEP). Burda and Mertens (2001) confirm that high-wage job losers in Germany can experience very large and persistent earnings losses. They also find that job losers in the bottom quartile of the wage distribution tend to exhibit significant gains from job loss.10 Our estimates are a larger and more persistent than a recent study of plant closings in Germany using the same administrative data source (Bender et al. 2002). These differences may arise partly due to differences in the definition of job loss, the estimation methodology, the sample used, and the time period covered. In particular, because we impose thirty days in unemployment to identify displaced workers, our approach may lead us to partially overstate the effect of job displacement. Part of the differences may also be due to our focus on workers losing their job in car manufacturing. Because the car-manufacturing sector is typically a high-wage sector, part of the losses we observe are due to losses in the industry wage premium. Table 4.3 also shows corresponding estimates for workers that had five years of predisplacement tenure and for workers exiting the large and stable firms analyzed in the first part of the chapter. As expected, highertenure workers experience larger and more persistent wage losses. However, the wage losses of workers leaving large firms are much larger. As found in von Wachter and Bender (2006), workers leaving large firms per10. Burda and Mertens’ (2001) estimates imply lower increases at the bottom and higher losses at the top. They do not focus on past starting wages, however, and have a somewhat different definition of layoff. For the top, they demonstrate that including recalls, as we do here, may underestimate the effect of job loss.

156

Till von Wachter and Stefan Bender

manently lose rents associated with jobs at large employers and never fully recover from the initial wage loss.11 The remainder of the section analyzes job displacement effects by previous starting wages on wage loss and post-loss wage levels. We obtain four key results: 1. There appears to be mean reversion. In particular, we find large differences in the degree of wage loss by previous starting wages, with the bottom gaining and the top losing. 2. There are permanent winners and losers from job loss. Those workers with high past starting wages experience permanent losses, whereas those with low starting wages experience long-term gains. 3. Pre-job loss starting wage and job tenure have a significant positive impact on wage levels after job loss. As expected, there is positive selection into high tenure and high past starting wages. 4. The effect of pre-job loss tenure and starting wage partly fades with time since job loss. These variables appear to serve as initial signal to the market of worker quality after a job loss. Table 4.3 and figure 4.9 show the estimates of percent wages lost at job loss for workers with high, medium, and low starting wages (based on the interquartile range of log real starting wages at the previous job). Clearly, workers in the high and medium starting wage groups suffer large and persistent losses and seem to drive the overall effect shown in figure 4.8. Thereby, the medium group tends to recover after ten years, whereas workers in the high group suffer permanent earnings losses of more than 10 percent. The group of workers with the smallest starting wages, on the other hand, has substantial benefits from the job loss that increase over time. These results suggest that starting wages contain firm-specific rents that fade upon job loss. The results also suggest that there is mean reversion in the labor market. This is consistent with a model of job search in which at job loss workers come from a different part of the wage distribution, but after job loss they are again reset to the mean of the wage distribution irrespective of their previous position. Note that we would not expect to see the effect of previous wages fade fully, as they are likely to contain some information on workers’ ability even beyond a person-fixed effect. 4.5.3 The Determinants of Post-Job Loss Wage Levels To explore this aspect further, table 4.4 analyzes the effect of pre-job loss characteristics on the level of log real wages after job loss. Thereby, the fo11. Large firms appear to provide an exceptional career environment that is permanently lost upon job displacement because, on average, workers will transit to a smaller firm. Von Wachter and Bender (2006) show that only apprentices who get displaced from large training firms suffer permanent losses in earnings relative to workers staying at the firm at the end of training. Once they control for the change in firm size at job loss, this excess loss disappears.

Do Initial Conditions Persist between Firms?

Fig. 4.9

157

Wage loss at job loss by starting wage at previous job

cus is particularly on the change in the effect of these characteristics over time, as this may further help discern the sources of persistence in the effect of initial conditions. We first replicate Kletzer’s (1989) basic model that includes previous job tenure as basic additional control in a standard human capital model of log wages. We confirm Kletzer’s result that past job tenure has a positive effect on current wage levels; in fact, despite the different definition of job loss, our point estimates are quite similar to hers. As in her case, this suggests that the positive correlation of tenure and wages not only arises from specific skills, but also from the fact that high-tenured workers are likely to be more able workers. The next column in table 4.4 also adds the log of previous starting wages to the Kletzer’s regression model. Again, we would expect past wages to have a positive effect on current wages as they are a function of components of workers’ skills not captured by observable characteristics. This is what we find—a 15 percent difference in starting wages raises wages past layoff by about 1 percent. In addition to being correlated with actual worker skills, part of the initial effect of past job tenure or past wages may be only temporary. To address this question, the last two columns of table 4.4 show estimates from regression models that interact characteristics of the past job with time since job displacement. Column (3) shows the estimates for past job tenure. When the interaction with past job loss is included, the initial effect doubles, and there is a clear pattern of decay. Thus, the estimates in column (1) capture the average effect of past job tenure all the years prior to job loss

158

Till von Wachter and Stefan Bender

Table 4.4

Effect of characteristics of lost job on wage levels after job displacement, three years pre-job loss tenure (1)

(2)

0.031 (0.0012) –0.0010 (0.0001) 0.0119 (0.0019) –0.00059 (0.0001) 0.043 (0.0010) 0.084 (0.0087) –0.00622 (0.00063) 0.000184 (0.000019) –0.000002 (0.00000)

0.066 (0.0068) 0.032 (0.0012) –0.0010 (0.0001) 0.0121 (0.0019) –0.00057 (0.0001) 0.041 (0.0010) 0.080 (0.0088) –0.00603 (0.00063) 0.000179 (0.000019) –0.000002 (0.00000)

231,185 0.36

231,185 0.36

Log starting wage at lost job (STWAGE) Years since displacement (YRSINCE) YRSINCE2 Tenure on past job (TEN) TEN2 Years of education (ED) Potential labor market experience (EXP) EXP2 EXP3 EXP4 TEN * YRSINCE ED * YRSINCE

(3)

(4)

0.063 (0.0025) –0.0009 (0.0001) 0.0217 (0.0021) –0.00074 (0.0001) 0.052 (0.0013) 0.082 (0.0089) –0.00626 (0.00064) 0.000189 (0.000019) –0.000002 (0.00000) –0.00127 (0.0001) –0.00186 (0.0002)

0.119 (0.0091) 0.100 (0.0056) –0.0009 (0.0001) 0.0223 (0.0021) –0.00074 (0.0001) 0.048 (0.0014) 0.074 (0.0090) –0.00582 (0.00065) 0.000179 (0.000019) –0.000002 (0.00000) –0.00129 (0.0001) –0.00148 (0.0002) –0.00959 (0.0013)

231,185 0.36

231,185 0.36

STWAGE * YRSINCE No. of observations R2

Notes: The entries in the table are coefficient estimates of regressions of log real daily wages after a job loss on characteristics of the lost job, year fixed effects, as well as individual characteristics. The specifications mirror closely that of Kletzer (1989). Regressors not listed in the table are a dummy for female and non-German, as well as nine dummies for industry, five dummies for occupation, a dummy for parttime status, and three dummies for blue- and white-collar status, all pertaining to the lost job. The regressions only include the first ten years after a job loss. Apprentices are excluded from the sample. Standard errors clustered at the individual level are in parentheses.

and obscure the fact that the effect fades over time. However, the effect does not fade completely even after ten years after job loss, suggesting, perhaps not surprisingly, that there is still an important correlation between past job tenure and unobserved worker skill. A similar pattern holds when past starting wages and their interaction with time since job loss are included in the model. The effect of past wages is initially larger and shows a linear pattern of decay (the estimates were not improved by including interactions with higher order polynomials of time since displacement). Again, the effect does not completely fade, suggesting

Do Initial Conditions Persist between Firms?

159

that conditional on observable characteristics past wages do contain information on workers’ productivity. However, after ten years, over 80 percent of the initial effect is gone. These results appear to be consistent with the hypothesis that the initial effect of past tenure and past starting wages captures temporary increases in reservation wages. Over time, reservation wages are determined by current market conditions, and the effect of past rents fades. In addition, as workers continue searching for jobs, their wage is again determined by their skills and overall wage distribution. Overall, we find that past starting wages contain firm-specific components of earnings that are partly lost when workers are displaced. These components may contain both group-level effects, such as firm-entry cohort effects or average firm-wage premiums, as well as individual-specific rents, for example, from job search. In future work, we plan to use displaced workers from a larger sample to distinguish between these different components. We also find that not all of the effect of past earnings is lost immediately. Some of the past wage may affect reservation wages and search efforts and fades only slowly over time as workers continue to search for jobs. Concluding, the benefit of getting a high paying job is mostly relegated to that job, but has positive spillover effects to future jobs that persist for up to ten years past a job loss. 4.6 Summary and Conclusion Persistence of entry conditions within firms has intrigued economists for a long time, but few studies were able to provide comprehensive empirical evidence on the incidence and causes of such cohort effects. In this chapter, we have used administrative information on wages and career patterns for all workers who ever worked in the German car industry matched to information on their establishments to make two contributions to the literature. First, we describe the incidence and size of firm-entry cohort effects for a large sample of firms. This allows us to study both the heterogeneity of cohort effects across our industry as well as their persistence both within and between firms. Second, we have begun to analyze the sources of persistent wage differences between different entry cohorts within firms. In particular, we have analyzed whether initial wage advantages are lost when workers lose their job and spend some time in unemployment. If initial wage differences are driven by differences in general human capital, they should persist when workers are forced to move to new jobs. If they are driven by firm-specific wage components, initial advantages should be lost at a job loss. To probe the degree of persistence of characteristics on the previous job further, we also analyzed the effect of past job tenure and previous starting wages on the level of wages after the job loss.

160

Till von Wachter and Stefan Bender

We find that firm-entry cohort effects are a common phenomenon among larger and stable firms in the German car manufacturing industry. Similar firms hiring similar workers at different points in time pay them different wages. We also find that these differences are quite heterogeneous among firms, such as they cannot be solely explained by marketwide business conditions. Initial wage differences between cohorts do tend to fade over time within firms. Similarly, firms’ wages tend to converge to a market wage. However, reversion of initial wage differences occurs slowly. In the second part, we find that initial wage differences are partly lost at job loss—high-wage workers have much larger and highly persistent wage losses. This suggests that wage differences prior to a job loss are in part driven by temporary firm-specific rents. Part of these rents is likely to consist of firm-entry cohort differences, but they may also contain workerspecific components such as search rents. We also observe mean reversion, that is, low-wage workers seem to permanently benefit from job loss. Consistent with the presence of temporary firm-specific wage components, past starting wages have an initial positive effect on wage levels after a job loss that fades over time. Pre-job loss characteristics appear to affect displaced workers’ reservation wages until their wage is again determined by their skills and the overall wage distribution. The results in this chapter highlight several important questions and areas for future research. First, it will be important to confirm our results with a wider sample of firms covering the entire German economy. An additional important question for future research is to establish to what extent worker mobility contributes to the reversion of initial differences in wages between entry cohorts. Third, using a larger sample we will be able to study the effect of exogenous events such as a mass layoff at the establishment level. Similarly, we will be able to distinguish the loss of groupspecific rents, such as average firm wage effects or cohort effects, from the loss of individual specific wage components arising among others from job search.

References Abowd, John, Patrick Corbel, and Francis Kramarz. 1999. The entry and exit of workers and the growth of employment: An analysis of French establishments. Review of Economics and Statistics 81(2): 170–87. Abowd, John M., and Francis Kramarz. 1999. The analysis of labor markets using matched employer-employee data. In Handbook of labor economics. Vol. 3B ed. O. Ashenfelter and D. Card, 2629–2710. Abraham, Katharine, and Henry Farber. 1987. Job duration, seniority, and earnings. American Economic Review 77 (3): 278–97.

Do Initial Conditions Persist between Firms?

161

Altonji, Joseph G., and Charles Pierret. 2001. Employer learning and statistical discrimination. Quarterly Journal of Economics 116 (1): 313–50. Baker, George, and Michael Gibbs, and Bengt Holmstrom. 1994. The wage policy of a firm. Quarterly Journal of Economics 109 (4): 881–919. Beaudry, Paul, and John DiNardo. 1991. The effect of implicit contracts and the movement of wages over the business cycle. Journal of Political Economy 99 (4): 665–88. Bender, Stefan, Christian Dustmann, David Margolis, and Costas Meghir. 2002. Worker displacement in France and Germany. In Losing work: International perspectives on job displacement, ed. Peter Kuhn, 375–470. Kalamazoo, MI: Upjohn Institute. Bender, Stefan, Annette Haas, and Christoph Klose. 2000. The IAB Employment Subsample 1975–1995. Journal of Applied Social Science Studies 120 (4): 649–62. Bender, Stefan, and Juergen Hilzendegen, Gotz Rohwer, and Helmut Rudolph. 1996. Die IAB-Beschaftigtenstichprobe 1975–1990: Eine praktische einführung (The IAB Employees Sample 1975–1990: A practical introduction). Beitraege zur Arbeitsmarkt- und Berufsforschung 197. Burda, Michael, and Antje A. Mertens. 2001. Estimating Wage Losses of Displaced Workers in Germany. Labor Economics 8 (1): 15–41. Couch, Kenneth A. 2001. Earnings losses and unemployment of displaced workers in Germany. Industrial and Labor Relations Review 54 (3): 559–72. Davis, Steven J., and John Haltiwanger. 1992. Gross job creation, gross job destruction, and employment reallocation. Quarterly Journal of Economics 107 (3): 819–63. Farber, Henry. 1997. The changing face of job loss in the United States, 1981–1995. Brookings Papers on Economic Activity, Microeconomics: 55–128. ———. 1999. Mobility and stability: The dynamics of job change in labor markets. In Handbook of labor economics. Vol. 3A, ed. O. Ashenfelter and D. Card, 2439– 83. Amsterdam: North Holland. ———. 2003. Job loss in the United States, 1981–2001. NBER Working Paper no. 9707. Cambridge, MA: National Bureau of Economic Research. Farber, Henry, and Robert Gibbons. 1996. Learning and wage dynamics. Quarterly Journal of Economics 111 (4): 1007–47. Gibbons, Robert, and Lawrence Katz. 1991. Layoffs and lemons. Journal of Labor Economics 9 (4): 351–80. Gibbons, Robert, and Michael Waldman. 2004. Task-specific human capital. AEA Papers and Proceedings 94:203–7. Görlitz, Katja, and Till von Wachter. 2006. The impact of job mobility on wage development: What role plays intermediate unemployment? RWI Essen. Mimeograph. Jacobson, Louis, Robert LaLonde, and Daniel Sullivan. 1993. Earnings losses of displaced workers. American Economic Review 83 (4): 685–709. Kahn, Lisa. 2006. The long-term labor market consequences of graduating college in a bad economy. Harvard University. Mimeograph. Kletzer, Lori G. 1989. Returns to seniority after permanent job loss. American Economic Review 79 (3): 536–43. Margolis, David. 1995. Firm heterogeneity and worker self-selection bias estimated returns to seniority. CIRANO Working Paper no. 95s-4. Montreal: Center for Interuniversity Research and Analysis on Organizations. Neal, Derek. 1995. Industry-specific human capital: Evidence from displaced workers. Journal of Labor Economics 13 (4): 653–77. Oi, Walter Y., and Todd L. Idson. 1999. Firm size and wages. In Handbook of labor

162

Till von Wachter and Stefan Bender

economics. Vol. 3B, ed. O. Ashenfelter and D. Card, 2165–2214. Amsterdam, North Holland. Okun, Arthur M. 1973. Upward mobility in a high-pressure economy. Brookings Papers of Economic Activity, issue no. 1:207–52. Washington, DC: Brookings Institution. Oreopoulos, Philip, Till von Wachter, and Andrew Heisz. 2006. Short- and longterm career effects of graduating in a recession: Hysteresis and heterogeneity in the market for college graduates. NBER Working Paper no. 12159. Cambridge, MA: National Bureau of Economic Research. Oyer, Paul. 2006. The macro-foundations of microeconomics: Initial labor market conditions and long-term outcomes for economists. NBER Working Paper no. 12157. Cambridge, MA: National Bureau of Economic Research. Parent, Daniel. 2000. Industry-specific capital and the wage profile: Evidence from the NLSY and the PSID. Journal of Labor Economics 18 (2): 306–23. Ruhm, Christopher. 1991. Are workers permanently scarred by job displacements? American Economic Review 81 (1): 319–24. Von Wachter, Till, and Stefan Bender. 2006. In the right place at the wrong time: The role of firms and luck in young workers’ careers. The American Economic Review 96 (5): 1679–1705.

5 Changes in Workplace Segregation in the United States between 1990 and 2000 Evidence from Matched Employer-Employee Data Judith Hellerstein, David Neumark, and Melissa McInerney

5.1 Introduction In recent work, we have constructed and described the 1990 Decennial Employer-Employee Dataset (DEED) based on matching records in the 1990 Decennial Census of Population to a Census Bureau list of most business establishments in the United States. We have used the 1990 DEED to estimate earnings and productivity differentials in manufacturing by demographic and skill group (Hellerstein and Neumark 2007), to study the influence of language skills on workplace segregation and wages (Hellerstein and Neumark 2003), to document the extent of workplace segregation by race and ethnicity, and to assess the contribution of residential segregation as well as skill to this segregation (Hellerstein and Neumark, forthcoming). We just recently completed the construction of the 2000 Beta-DEED Judith Hellerstein is an associate professor of economics at the University of Maryland, and a research associate of the National Bureau of Economic Research. David Neumark is a professor of economics at the University of California, Irvine, a research fellow of the Institute for the Study of Labor, and a research associate of the National Bureau of Economic Research. Melissa McInerney is a statisician at the U.S. Bureau of the Census, Center for Economic Studies, and a PhD candidate at the University of Maryland, Department of Economics. This research was funded by National Institute of Child Health & Human Development (NICHD) grant R01HD042806. We also thank the Alfred P. Sloan Foundation for its generous support. We are grateful to Ron Jarmin, Julia Lane, and an anonymous reviewer for helpful comments. The analysis and results presented in this paper are attributable to the authors and do not necessarily reflect concurrence by the Center for Economics Studies, the U.S. Bureau of the Census, or the Sloan Foundation. This paper has undergone a more limited review by the Census Bureau than its official publications. It has been screened to ensure that no confidential data are revealed.

163

164

Judith Hellerstein, David Neumark, and Melissa McInerney

(based on the 2000 Census of Population).1 In this paper, we use the 1990 and 2000 DEEDs to measure changes in establishment-level workplace segregation over the intervening decade, an analysis for which the DEEDs are uniquely well-suited. We study segregation by education, by race and Hispanic ethnicity, and by sex. With respect to segregation by race and ethnicity, this work is complementary to a flurry of research studying changes in residential segregation from 1990 to 2000 (Glaeser and Vigdor 2001; Iceland and Weinberg 2002; and McConville and Ong 2001). As we have suggested elsewhere (and see Estlund 2003), however, workplace segregation may be far more salient for interactions between racial and ethnic groups than is residential segregation. The boundaries used in studying residential segregation may not capture social interactions and are to some extent explicitly drawn to accentuate segregation among different groups; for example, Census tract boundaries are often generated in order to ensure that the tracts are “as homogeneous as possible with respect to population characteristics, economic status, and living conditions.”2 In contrast, workplaces—specifically establishments—are units of observation that are generated by economic forces and in which people clearly do interact in a variety of ways, including work, social activity, labor market networks, and so on. Thus, while it is more difficult to study workplace segregation because of data constraints, measuring workplace segregation may be more useful than measuring residential segregation, as traditionally defined, for describing the interactions that arise in society between different groups in the population.3 Of course, similar arguments to those about workplaces could be made about other settings, such as schools, religious institutions, and so on (e.g., James and Taeuber 1985), but data constraints truly prevent saying much of anything about segregation along these lines. Segregation is potentially important for a number of reasons. Aside from general social issues regarding integration between different groups, labor market segregation by race and ethnicity accounts—at least in a statistical sense—for a sizable share of wage gaps between white males and other demographic groups (e.g., Carrington and Troske 1998a; Bayard et al. 1999; King 1992; Watts 1995; Higgs 1977), and the same is true of labor market segregation by sex (Bayard et al. 2003; Blau 1977; and Groshen 1. The 2000 Beta-DEED is an internal U.S. Census Bureau data set that will ultimately become part of an integrated matched employer-employee database at the U.S. Census Bureau. The new integrated data will have characteristics of the Decennial Employer-Employee Database (DEED) and the Longitudinal Employer-Household Dynamics Program (LEHD). Hereafter, the 2000 Beta-DEED will be referred to as the 2000 DEED. 2. See the U.S. Census Bureau, http://www.census.gov/geo/www/GARM/Ch10GARM.pdf (viewed April 27, 2005). Echenique and Fryer (2005) develop a segregation index that relies much less heavily on ad hoc definitions of geographical boundaries. 3. Moreover, industry code, the closest proxy in public-use data to an establishment identifier, is a very crude measure to use to examine segregation. For example, we calculate that racial and ethnic segregation at the three-digit industry level in the DEED is typically on the order of one-third as large as the establishment-level segregation we document in the following.

Changes in Workplace Segregation in the United States

165

1991).4 There has generally been less attention paid to segregation by education, but in our earlier work (Hellerstein and Neumark, forthcoming), we documented rather extensive segregation by education (as well as language, which we do not consider in the present paper) in the 1990 DEED. Measuring changes in workplace segregation along these lines is of interest for a number of reasons. First, although much attention has been paid to changes in residential segregation—of which there is evidence of modest declines from 1990 to 2000—changes in workplace segregation may be more salient to understanding changing social forces. Second, aside from the relative importance of workplace and residential segregation, in the United States there are extensive efforts to reduce labor market discrimination, and, therefore, measuring changes in workplace segregation by race, ethnicity, and sex provides indicators of the success of these efforts. Finally, increases in the productivity (and pay) of more-educated workers relative to less-educated workers may have led to increased segregation by skill (e.g., Kremer and Maskin 1996).5 A comparison of education segregation between 1990 and 2000 possibly can shed some light on this hypothesis although relatively more of the run-up in wage inequality occurred prior to 1990 (Autor, Katz, and Kearney 2005). We measure changes in segregation using the 1990 and 2000 Decennial Employer-Employee Databases (DEEDs). For each year, the DEED is based on matching records in the Decennial Census of Population for that year to a Census Bureau list of most business establishments in the United States. The matching yields data on multiple workers matched to establishments, providing the means to measure workplace segregation (and changes therein) in the United States based on a large, fairly representative data set. In addition, the data from the Decennial Census of Population provides the necessary information on race, ethnicity, and so on. Thus, data from the 1990 and 2000 DEEDs provides unparalleled opportunities to study changes in workplace segregation by skill, race, ethnicity, and sex.6 4. This segregation may occur along industry and occupation lines, as well as at the more detailed level of the establishment or job cell (occupations within establishments). For example, Bayard et al. (1999) found that, for men, job-cell segregation by race accounts for about half of the black-white wage gap and a larger share of the Hispanic-white wage gap. 5. For example, let the production function be f (L1, L2)  L1cL2d, with d  c. Assume that there are two types of workers: unskilled workers (L1) with labor input equal to one efficiency unit, and skilled workers (L2) with efficiency units of q  1. Kremer and Maskin (1996) show that for low q, it is optimal for unskilled and skilled workers to work together, but above a certain threshold of q (that is, a certain amount of skill inequality), the equilibrium will reverse, and workers will be sorted across firms according to skill. Thus, as the returns to education rise (q increases), there may be increased segregation by education. 6. Carrington and Troske (1998a, b) use data sets much more limited in scope than the ones we use here to examine workplace segregation by race and sex. In general, the paucity of research on workplace segregation is presumably a function of the lack of data linking workers to establishments.

166

Judith Hellerstein, David Neumark, and Melissa McInerney

5.2 The 1990 and 2000 DEEDs The analysis in this paper is based on the 1990 and 2000 DEEDs, which we have created at the Center for Economic Studies at the U.S. Bureau of the Census. We have described the construction of the 1990 DEED in detail elsewhere (in particular, Hellerstein and Neumark 2003). The construction of the 2000 DEED follows the same procedures, and our detailed investigation of the 2000 data thus far has indicated that no new serious problems arise that require different methods for 2000. Thus, in this section we simply provide a quick overview of the construction of the data sets. The DEED for each year is formed by matching workers to establishments. The workers are drawn from the Sample Edited Detail File (SEDF), which contains all individual responses to the Decennial Census of Population one-in-six Long Form. The establishments are drawn from the Census Bureau’s Business Register list (BR), formerly known as the Standard Statistical Establishment List (SSEL); the BR is a database containing information for most business establishments operating in the United States in each year, which is continuously updated (see Jarmin and Miranda 2002). Households receiving the Decennial Census Long Form were asked to report the name and address of the employer in the previous week for each employed member of the household. The file containing this employer name and address information is referred to as the “Write-In” file, which contains the information written on the questionnaires by Long-Form respondents but not actually captured in the SEDF. The BR is a list of most business establishments with one or more employees operating in the United States. The Census Bureau uses the BR as a sampling frame for its Economic Censuses and Surveys and continuously updates the information it contains. The BR contains the name and address of each establishment, geographic codes based on its location, its four-digit Standard Industrial Classification (SIC) code, and an identifier that allows the establishment to be linked to other establishments that are part of the same enterprise and to other Census Bureau establishment- or firm-level data sets that contain more detailed employer characteristics. We can, therefore, use employer names and addresses for each worker in the Write-In file to match the Write-In file to the BR. Because the name and address information on the Write-In file is also available for virtually all employers in the BR, nearly all of the establishments in the BR that are classified as “active” by the Census Bureau are available for matching. Finally, because both the Write-In file and the SEDF contain identical sets of unique individual identifiers, we can use these identifiers to link the Write-In file to the SEDF. Thus, this procedure yields a very large data set with workers matched to their establishments, along with all of the information on workers from the SEDF. Matching workers and establishments is a difficult task because we would not expect employers’ names and addresses to be recorded identi-

Changes in Workplace Segregation in the United States

167

cally on the two files. To match workers and establishments based on the Write-In file, we use MatchWare—a specialized record linkage program. MatchWare is comprised of two parts: a name and address standardization mechanism (AutoStan) and a matching system (AutoMatch). This software has been used previously to link various Census Bureau data sets (Foster, Haltiwanger, and Krizan 1998). Our method to link records using MatchWare involves two basic steps. The first step is to use AutoStan to standardize employer names and addresses across the Write-In file and the BR. Standardization of addresses in the establishment and worker files helps to eliminate differences in how data are reported. The standardization software considers a wide variety of different ways that common address and business terms can be written and converts each to a single standard form. Once the software standardizes the business names and addresses, each item is parsed into components. The value of parsing the addresses into multiple pieces is that we can match on various combinations of these components. We supplemented the AutoStan software by creating an acronym for each company name and added this variable to the list of matching components.7 The second step of the matching process is to select and implement the matching specifications. The AutoMatch software uses a probabilistic matching algorithm that accounts for missing information, misspellings, and even inaccurate information. This software also permits users to control which matching variables to use, how heavily to weight each matching variable, and how similar two addresses must be in order to constitute a match. AutoMatch is designed to compare match criteria in a succession of “passes” through the data. Each pass is comprised of “Block” and “Match” statements. The Block statements list the variables that must match exactly in that pass in order for a record pair to be linked. In each pass, a worker record from the Write-In file is a candidate for linkage only if the Block variables agree completely with the set of designated Block variables on analogous establishment records in the BR. The Match statements contain a set of additional variables from each record to be compared. These variables need not agree completely for records to be linked, but are assigned weights based on their value and reliability. For example, we might assign “employer name” and “city name” as Block variables and assign “street name” and “house number” as Match variables. In this case, AutoMatch compares a worker record only to those establishment records with the same employer name and city name. All employer records meeting these criteria are then weighted by whether and 7. For 2000, we also added standard acronyms or abbreviations for cities, such as NY or NYC and LA. However, this added a negligible number of additional matches, so we did not go back and do the same for the 1990 DEED.

168

Judith Hellerstein, David Neumark, and Melissa McInerney

how closely they agree with the worker record on the street name and house number Match specifications. The algorithm applies greater weights to items that appear infrequently. The employer record with the highest weight will be linked to the worker record conditional on the weight being above some chosen minimum. Worker records that cannot be matched to employer records based on the Block and Match criteria are considered residuals, and we attempt to match these records on subsequent passes using different criteria. It is clear that different Block and Match specifications may produce different sets of matches. Matching criteria should be broad enough to cover as many potential matches as possible, but narrow enough to ensure that only matches that are correct with a high probability are linked.8 Because the AutoMatch algorithm is not exact, there is always a range of quality of matches, and we, therefore, are cautious in accepting linked record pairs. Our general strategy is to impose the most stringent criteria in the earliest passes and to loosen the criteria in subsequent passes, while always maintaining criteria that err on the side of avoiding false matches. We choose matching algorithms based on substantial experimentation and visual inspection of many thousands of records. The final result is an extremely large data set, for each year, of workers matched to their establishment of employment. The 1990 DEED consists of information on 3.29 million workers matched to around 972,000 establishments, accounting for 27.1 percent of workers in the SEDF and 18.6 percent of establishments in the BR. The 2000 DEED consists of information on 4.09 million workers matched to around 1.28 million establishments, accounting for 29.1 percent of workers in the SEDF and 22.6 percent of establishments in the BR.9 In table 5.1, we provide descriptive statistics for the matched workers from the DEED as compared to the SEDF. Columns (1) and (4) report summary statistics for the SEDF for the sample of workers who were elig8. One might also considering trying to impute matches where this strategy fails by matching based on imputed place of work instead of information in the Write-In file. However, this turns out to be problematic. Even imputing place of work at the level of the Census tract is not easy. For example, there are workers in the SEDF that we are able to match to an employer in the DEED using name and address information whose place of work code actually is allocated in the SEDF. For these workers, the allocated Census tract in the SEDF disagrees with the BR Census tract of the matched establishment in more than half the cases. 9. For both the DEED and SEDF, we have excluded individuals as follows: with missing wages; who did not work in the year prior to the survey year or in the reference week for the Long Form of the Census; who did not report positive hourly wages; who did not work in one of the fifty states or the District of Columbia (whether or not the place of work was imputed); who were self-employed; who were not classified in a state of residence; or who were employed in an industry that was considered “out-of-scope” in the BR. (Out-of-scope industries do not fall under the purview of Census Bureau surveys. They include many agricultural industries, urban transit, the U.S. Postal Service, private households, schools and universities, labor unions, religious and membership organizations, and government/public administration. The Census Bureau does not validate the quality of BR data for businesses in out-of-scope industries.)

Changes in Workplace Segregation in the United States Table 5.1

Means for workers 1990

Age Female Married White Hispanic Black Full-time No. of kids (if female) High school diploma Some college BA Advanced degree Ln(hourly wage) Hourly wage Hours worked in previous year Weeks worked in previous year Earnings in previous year Industry Mining Construction Manufacturing Transportation Wholesale Retail FIRE Services N

169

2000

SEDF (1)

Full DEED (2)

Restricted DEED (3)

SEDF (4)

Full DEED (5)

Restricted DEED (6)

37.08 (12.78) 0.46 0.60 0.82 0.07 0.08 0.77 0.75 (1.04) 0.34 0.30 0.13 0.05 2.21 (0.70) 12.10 (82.19)

37.51 (12.23) 0.47 0.65 0.86 0.05 0.05 0.83 0.73 (1.01) 0.33 0.32 0.16 0.05 2.30 (0.65) 12.89 (37.07)

37.53 (12.13) 0.47 0.63 0.84 0.06 0.06 0.84 0.69 (0.99) 0.30 0.33 0.18 0.06 2.37 (0.65) 13.68 (27.41)

39.15 (13.03) 0.46 0.58 0.78 0.09 0.09 0.78 0.78 (1.07) 0.31 0.33 0.15 0.06 2.55 (0.73) 17.91 (137.20)

39.57 (12.51) 0.50 0.62 0.83 0.07 0.06 0.82 0.76 (1.04) 0.29 0.35 0.18 0.08 2.63 (0.70) 18.83 (63.61)

39.53 (12.33) 0.51 0.60 0.79 0.08 0.08 0.83 0.74 (1.03) 0.25 0.35 0.20 0.09 2.70 (0.70) 20.19 (64.05)

39.51 (11.44)

40.42 (10.37)

40.55 (10.10)

40.22 (11.74)

40.72 (11.09)

40.90 (10.85)

46.67 (11.05) 22,575 (26,760)

48.21 (9.34) 25,581 (29,475)

48.46 (9.05) 27,478 (30,887)

47.23 (10.58) 33,521 (42,977)

48.38 (9.27) 37,244 (47,237)

48.56 (9.05) 40,272 (50,406)

0.01 0.07 0.25 0.08 0.05 0.20 0.08 0.26

0.01 0.04 0.34 0.05 0.07 0.17 0.08 0.24

0.01 0.03 0.35 0.05 0.08 0.15 0.09 0.24

0.00 0.05 0.26 0.05 0.05 0.21 0.07 0.31

0.00 0.04 0.26 0.05 0.05 0.20 0.07 0.32

12,143,183

3,291,213

1,828,020

4,089,098

2,209,908

0.01 0.08 0.21 0.07 0.05 0.21 0.07 0.31 14,057,121

ible to be matched to their establishments, for 1990 and 2000, respectively. Columns (2) and (5) report summary statistics for the full DEED sample. For both years, the means of the demographic variables in the full DEED are quite close to the means in the SEDF across most dimensions. For example, for the 1990 data, female workers comprise 46 percent of the SEDF

170

Judith Hellerstein, David Neumark, and Melissa McInerney

and 47 percent of the full DEED, and the number of children (for women) is 0.75 in the SEDF and 0.73 in the DEED. Nonetheless, there are cases of somewhat larger differences. Race and ethnic differences are larger in both years; for example, in 2000, the percent white is 78 in the SEDF versus 83 in the DEED, and, correspondingly, the share black (and also Hispanic) is lower in the DEED. In addition, the percent female in the 2000 data is 46 in the SEDF, but 50 in the DEED; this is different than the discrepancy in 1990 where the percent female is 46 in the SEDF and only a slightly higher 47 percent in the DEED. Part of the explanation for differences in racial and ethnic representation that result from the matching process is that there are many individuals who meet our sample inclusion criteria but for whom the quality of the business address information in the Write-In file is poor, and race and ethnic differences in reporting account for part of the differences in representation. We suspect that the differences in business address information partially reflect weaker labor market attachment among minorities, suggesting that the segregation results we obtain might best be interpreted as measuring the extent of segregation among workers who have relatively high labor force attachment and high attachment to their employers. The last eight rows of the table report on the industry distribution of workers. We do find some overrepresentation of workers in manufacturing—more so in 1990 when manufacturing comprised a larger fraction of workers to begin with in the SEDF. The reasons for this are given in the following when we discuss establishment-level data. Columns (3) and (6) report summary statistics for the workers in the DEED who comprise the sample from which we calculate segregation measures. The sample size reductions relative to columns (2) and (5) arise for two reasons. First, for reasons explained in the methods section, we exclude workers who do not live and work in the same Metropolitan Statistical Area/Primary Metropolitan Statistical Area (MSA/PMSA). Second, we exclude workers who are the only workers matched to their establishments, as there are methodological advantages to studying segregation in establishments where we observe at least two workers. The latter restriction effectively causes us to restrict the sample to workers in larger establishments, which is the main reason why some of the descriptive statistics are slightly different between the second and third columns (for example, slightly higher wages and earnings in columns [3] and [6]). In addition to comparing worker-based means, it is useful to examine the similarities across establishments in the BR and the DEED for each year. Table 5.2 shows descriptive statistics for establishments in each data set. As column (1) indicates, there are 5,237,592 establishments in the 1990 BR, and of these 972,436 (18.6 percent) also appear in the full DEED for 1990, as reported in column (2). For 2000, the percentage in the full DEED is somewhat higher (22.6). Because only one in six workers are sent De-

Changes in Workplace Segregation in the United States Table 5.2

Means for establishments 1990

BR (1) Total employment Establishment size 1–25 26–50 51–100 101+ Industry Mining Construction Manufacturing Transportation Wholesale Retail FIRE Services In MSA Census region North East Mid Atlantic East North Central West North Central South Atlantic East South Central West South Central Mountain Pacific Payroll ($1,000) Payroll/total employment Share of employees matched Multiunit establishment N

171

17.57 (253.75)

2000

Full DEED (2)

Restricted DEED (3)

52.68 (577.39)

104.67 (996.52)

BR (4) 18.77 (138.11)

Full DEED (5) 48.74 (232.05)

Restricted DEED (6) 95.54 (371.18)

0.88 0.06 0.03 0.03

0.65 0.15 0.10 0.10

0.39 0.22 0.19 0.21

0.87 0.06 0.03 0.03

0.66 0.15 0.09 0.09

0.41 0.21 0.17 0.20

0.00 0.09 0.06 0.04 0.08 0.25 0.09 0.28 0.81

0.01 0.07 0.13 0.05 0.11 0.24 0.10 0.26 0.82

0.01 0.06 0.23 0.05 0.10 0.23 0.11 0.21 1

0.00 0.11 0.06 0.04 0.07 0.25 0.09 0.35 0.81

0.00 0.08 0.13 0.05 0.07 0.29 0.08 0.30 0.79

0.00 0.07 0.18 0.05 0.07 0.27 0.07 0.27 1

0.06 0.16 0.16 0.07 0.18 0.05 0.10 0.06 0.16 397 (5,064) 21.02 (1,385.12)

0.05 0.16 0.21 0.07 0.15 0.04 0.09 0.05 0.17 2,910 (16,601) 26.70 (181.48) 0.16 0.53

0.06 0.15 0.16 0.08 0.18 0.06 0.10 0.07 0.16 694.44 (69,383) 33.74 (772.29)

0.23

0.06 0.15 0.20 0.08 0.16 0.05 0.10 0.05 0.15 1,358 (10,329) 24.24 (111.79) 0.17 0.42

0.26

0.05 0.14 0.20 0.09 0.16 0.05 0.10 0.06 0.15 1,993 (115,076) 35.91 (1,834.40) 0.16 0.40

0.04 0.14 0.21 0.08 0.16 0.04 0.10 0.06 0.17 4,421 (198,414) 42.27 (1,877.29) 0.14 0.50

5,237,592

972,436

317,112

5,651,680

1,279,999

411,300

cennial Census Long Forms, it is more likely that large establishments will be included in the DEED. One can see evidence of the bias toward larger employers by comparing the means across data sets for total employment. (This bias presumably also influences the distribution of workers and establishments across industries, where, for example, the DEEDs overrepresent workers in manufacturing establishments.) On average, establishments in the BRs have eighteen to nineteen employees, while the average in

172

Judith Hellerstein, David Neumark, and Melissa McInerney

the DEEDs is forty-nine to fifty-three workers. The distributions of establishments across industries in the DEED relative to the BR are similar to those for workers in the worker sample. In columns (3) and (6), we report descriptive statistics for establishments in the restricted DEEDs, corresponding to the sample of workers in columns (3) and (6) of table 5.1. In general, the summary statistics are quite similar between columns (2) and (3) and between columns (5) and (6), with an unsurprising right shift in the size distribution of establishments. Overall, however, the DEED samples are far more representative than previous detailed matched data sets for the United States constructed using just the SEDF and the BR (see Hellerstein and Neumark 2003).10 Because the DEED captures larger establishments and because our sample restrictions accentuate this, our analysis focuses on larger establishments. So, for example, the first quartile of the establishment size distribution for workers in our analysis is approximately forty-one workers in 1990 and thirty-six in 2000, whereas the first quartile of the employmentweighted size distribution of all establishments in the BR for each year is nineteen in 1990 and twenty-one in 2001.11 Although we acknowledge that it would be nice to be able to measure segregation in all establishments, this is not the data set with which to do that convincingly. Nonetheless, most legislation aimed at combating discrimination is directed at larger establishments; Equal Employment Opportunity Commission (EEOC) laws cover employers with fifteen or more workers, and affirmative action rules for federal contractors cover employers with fifty or more workers. Because policy has been directed at larger establishments, examining the extent of and changes in workplace segregation in larger establishments is important. 5.3 Methods We focus our analysis on a measure of segregation that is based on the percentages of workers in an individual’s establishment, or workplace, in different demographic groups. Consider for clarity measuring segregation between white and Hispanic workers. For each white or Hispanic worker in our sample, we compute the percentage of Hispanic workers with which that worker works, excluding the worker him- or herself. Because we exclude 10. These earlier matched data sets—the Worker-Establishment Characteristics Database (WECD), which covers manufacturing only, and the New Worker-Establishment Characteristics Database (NWECD), which covers all industries—were smaller and less representative because the matching algorithm used could only be applied to establishments that were unique in a cell defined by detailed geographic information and industry classification. Thus, for example, manufacturing establishments were much more likely to occupy their own industry-location cell than were retail establishments. 11. In order to adhere to U.S. Census Bureau confidentiality rules, these are “pseudo quartiles” based on averages of observations symmetrically distributed around the actual quartiles.

Changes in Workplace Segregation in the United States

173

an individual’s own ethnicity in this calculation, our analysis of segregation is conducted on establishments where we observe at least two workers. We then average these percentages separately for white workers in our sample and for Hispanic workers. These averages are segregation measures commonly used in the sociology literature. The average percentage of coworkers in Hispanic workers’ establishments who are Hispanic, denoted HH, is called the “isolation index,” and the average percentage of coworkers in white workers’ establishments who are Hispanic, denoted WH, is called the “exposure index.” We focus more on a third measure, the difference between these, or CW  HH – WH, as a measure of “coworker segregation.” The variable CW measures the extent to which Hispanics are more likely than are whites to work with other Hispanics. For example, if Hispanics and whites are perfectly segregated, then HH equals 100, WH is zero, and CW equals 100.12 We first report observed segregation, which is simply the sample mean of the segregation measure across workers. We denote this measure by appending an O superscript to the coworker segregation measure—that is, CWO. One important point that is often overlooked in research on segregation, however, is that some segregation occurs even if workers are assigned randomly to establishments, and we are presumably most interested in the segregation that occurs systematically—that is, that which is greater than would be expected to result from randomness (Carrington and Troske 1997). Rather than considering all deviations from proportional representation across establishments as an “outcome” or “behavior” to be explained, we subtract from our measured segregation the segregation that would occur by chance if workers were distributed randomly across establishments, using Monte Carlo simulations to generate measures of randomly occurring segregation. We denote this random segregation CWR (and similarly for the isolation and exposure indexes) and then focus on the difference (CWO – CWR), which measures segregation above and beyond that which occurs randomly.13 Although theoretically one can have CWO  CWR (that is, there is less segregation than would be generated randomly) or CWO  CWR, only the latter occurs in practice in our data. Again following Carrington and Troske, we scale this difference by the maximum 12. We could equivalently define the percentages of white workers with which Hispanic or white workers work, HW and WW , which would simply be 100 minus these percentages, and CW  WW  HW. 13. This distinction between comparing measured segregation to a no-segregation ideal or segregation that is generated by randomness is discussed in other work (see, e.g., Cortese, Falk, and Cohen 1976; Winship 1977; Boisso et al. 1994; and Carrington and Troske 1997). Of course, to build CWR we also compute the isolation and exposure indexes that would be generated in the case of random allocation of workers, and we report these as well.

174

Judith Hellerstein, David Neumark, and Melissa McInerney

segregation that can occur, or (100 – CWR), we refer to this measure as “effective segregation.” Thus, the effective segregation measure is





CWO – CWR  100, 100 – CWR

which measures the share of the maximum possible segregation that is actually observed. There are two reasons that we exclude the worker’s own ethnicity when computing the fraction of Hispanics with which he or she works. First, this ensures that, in large samples of workers, if workers are randomly allocated across establishments, HH and WH both equal the share Hispanic in the population. That is, in the case of random allocation, we expect to have CWR equal to 0. This is a natural scaling to use and stands in contrast to what happens when the worker is included in the calculations, where CWR will exceed 0 because Hispanic workers are treated as working with “themselves.” Second, and perhaps more important, when the own worker is excluded, our segregation measures are invariant to the sizes of establishments studied. To see this in a couple of simple examples, first consider a simple case of an economy with equal numbers of Hispanics and whites all working in two-person establishments. Establishments can therefore be represented as HH (for two Hispanic workers), HW, or WW. With random allocation, 1/4 of establishments are HH, 1/2 are WH, and 1/4 are WW. Thus, excluding the own worker, HHR  (1/2)  1 (1/2)  0  1/2, WHR  (1/2)  1 (1/2)  0  1/2, and CWR  0.14 If we count the individual, then HHR  (1/2)  1 (1/2)  (1/2)  3/4, WHR  (1/2)  (1/2) (1/2)  0  1/4, and CWR  1/2. With three-worker establishments and random allocation, 1/8 of establishments are HHH (employing 1/4 of Hispanic workers), 1/8 are WWW (employing 1/4 of white workers), 3/8 are HWW (employing 1/4 of Hispanic and 1/2 of white workers), and 3/8 are HHW (employing 1/2 of Hispanic and 1/4 of white workers). Going through the same type of calculation as in the preceding, if we include the worker, then HHR  (1/4)  1 (1/4)  (1/3) (1/2)  (2/3)  2/3, WHR  (1/4)  0 (1/4)  (2/3) (1/2)  (1/3)  1/3 and CWR  1/3, whereas if we exclude the worker we again get HHR  1/2, WHR  1/2, and CWR  0. Although we just argued that in the case of random allocation Hispanics and whites should work with equal percentages of Hispanic coworkers on average (so that CWR is zero), this result may not hold in parts of our analysis for two reasons. First, this is a large-sample result, and although the baseline sample size in our data set is large, the samples that we use to calculate some of our segregation measures are not necessarily large enough to generate this asymptotic result. Second, some of our segregation 14. For the first calculation, for example, 1/2 of hispanic workers are in HH establishments, for which the share hispanic is 1, and 1/2 are in WH establishments, for which the share Hispanic (excluding the worker) is 0.

Changes in Workplace Segregation in the United States

175

measures are calculated conditional on geography (in particular, MSA/ PMSA of residence), for reasons explained in the following. When we condition on geography, we calculate the extent of segregation that would be expected if workers were randomly allocated across establishments within a geographic area. If Hispanics and whites are not evenly distributed across geographic borders, random allocation of workers within geographical areas still will yield the result that Hispanics are more likely to have Hispanic coworkers than are white workers because, for example, more Hispanics will come from areas where both whites and Hispanics work with a high share of Hispanic workers. For these reasons, in order to determine how much segregation would occur randomly, in all cases we conduct Monte Carlo simulations of the extent of segregation that would occur with random allocation of workers. There are, of course, other possible segregation measures, such as the traditional Duncan index (Duncan and Duncan 1955) or the Gini coefficient. We prefer the coworker segregation measure (CW) to these other measures for two reasons. First, the Duncan and Gini measures are scale invariant, meaning that they are insensitive to the proportions of each group in the workforce. For example, if the number of Hispanics doubles but they are allocated to establishments in the same proportion as the original distribution, the Duncan and Gini indexes are unchanged. However, except in establishments that are perfectly segregated, the doubling of Hispanics leads each Hispanic worker in the sample to work with a larger percentage of Hispanic coworkers and also each white worker to work with more Hispanics. In general, this implies that both the isolation and exposure indexes (HH and WH, respectively), will increase. But the isolation index will increase by more because establishments with more Hispanics to begin with will have larger increases in the number of Hispanic workers, and, hence, CW will increase.15 In our view, this kind of increase in the number of Hispanic workers should be characterized as an increase in segregation. Second, these alternative segregation measures are also sensitive to the number of matched workers in an establishment (the same issue outlined in the preceding), and because they are measures that are calculated at only the establishment level—unlike the coworker segregation measure we use—there is no conceptual parallel to excluding the own worker from the calculation.16 15. More generally, WH will also increase, but not by as much as HH, and CW will, therefore, rise. For perhaps the simplest such case, start with four establishments as follows: one HHH, one HHW, one HWW, and one WWW. In this case, HH  2/3, WH  1/3, and CW  1/3. Doubling the number of Hispanics and allocating them proportionally, we get the following four establishments: HHHHHH, HHHHW, WWHH, and WWW: In this case HH rises to 29/36 (increasing by 5/36), WH rises to 14/36 (increasing by 2/36), and CW rises to 15/36 (increasing by 3/36). 16. We believe this explains why, in Carrington and Troske (1998a, table 3), where there are small samples of workers within establishments, the random Gini indexes are often extremely high.

176

Judith Hellerstein, David Neumark, and Melissa McInerney

At the same time, because calculated changes in segregation between 1990 and 2000 based on our coworker segregation index are sensitive to the overall proportions of each group in the workforce, changes over the decade in the proportions of particular demographic groups that are matched to establishments can generate changes in measured segregation. So, for example, the fact that the fraction of workers who are Hispanic grew from 1990 to 2000 should yield a small increase in measured coworker segregation by ethnicity over the decade (even if Hispanics and whites are distributed across establishments in the same proportion in each year). We could avoid this problem by using scale-invariant segregation measures, but then we would fail to capture changes in segregation due to actual changes in workforce composition. That is, the fact that Hispanics make up a growing fraction of the workforce is an important phenomenon to capture.17 Nonetheless, although we emphasize the coworker segregation measure throughout, we also report our key results based on the Duncan index to see how robust the conclusions are. We present some “unconditional” nationwide segregation measures, as well as “conditional” measures that first condition on metropolitan area (MSA/PMSA) of residence. In the first, the simulations randomly assign workers to establishments anywhere in the country; not surprisingly, in these simulations the random segregation measures are zero or virtually indistinguishable from zero. For comparability, when we construct these unconditional segregation measures, we use only the workers included in the MSA/PMSA sample used for the conditional analysis.18 The unconditional estimates provide the simplest measures of the extent of integration by skill, race, ethnicity, or sex in the workplace. However, they reflect the distribution of workers both across cities and across establishments within cities. As such, the unconditional measures may tell us less about forces operating in the labor market to create segregation, whereas the conditional measures—which can be interpreted as taking residential segregation by city as given—may tell us more about these forces. Because we use the same samples for the conditional and unconditional analyses, for these analyses the observed segregation measures are identical. Only the simulations differ, but these differences, of course, imply differences in the effective segregation measures. 17. Some measured changes in the sample composition of workers over time may reflect changes in the match rates of various kinds of workers to establishments rather than a change in the underlying population composition. This is obviously a limitation of matched data sets like ours, one that exists to a much smaller extent in administrative data sets that come closer to capturing fully the universe of workers. 18. The results in this paper are generally robust to measuring unconditional segregation by including all workers in the United States whether they live and work in a metropolitan area. For the unconditional analysis using the full DEEDs versus the MSA/PMSA sample, the changes in segregation are always in the same direction and qualitatively similar although the estimated percentage changes are a bit more moderate than those reported in the following.

Changes in Workplace Segregation in the United States

177

For the Monte Carlo simulations that generate measures of random segregation, we need to first define the unit within which we are considering workers to be randomly allocated. This requires a specification of the relevant labor market. We use U.S. Census Bureau MSA/PMSA designations because these are defined to some extent based on areas within which substantial commuting to work occurs.19 An MSA is a set of one or more counties that contains a population center and the adjacent densely-settled counties, with additional counties included if the share of residents commuting to the population core exceeds a certain threshold.20 In the case of particularly large MSAs, such as Washington, DC-Baltimore, MD, the entire region meets the criteria to be a MSA, and two or more subsets of the region also meet the MSA definition. In cases such as these, we consider the smaller subsets of counties, called PMSAs. In the Washington, DCBaltimore, MD example, the larger area (called a Consolidated Metropolitan Statistical Area, or CMSA) is comprised of three PMSAs: Baltimore, MD; Hagerstown, MD; and Washington, DC. Thus, the metropolitan areas on which we focus should be relatively well-defined labor markets, rather than huge areas covering many cities.21 For example, the 10th percentile of the distribution of MSA/PMSA populations is comprised of smaller metropolitan areas such as Sheboygan, WI, with approximately 100,000 residents, and the 90th percentile is Sacramento, CA, having roughly 1.6 million residents.22 At the same time, we are certainly not claiming that residential segregation at a level below that of the MSA/PMSA does not influence workplace segregation. However, an analysis of this question requires somewhat different methods. For example, in conducting the simulations, it is not obvious how one should limit the set of establishments within a metropolitan area in which a worker could be employed. Returning to the simulation procedure, we calculate for each MSA/ 19. See the U.S. Census Bureau, http://www.census.gov/geo/lv4help/cengeoglos.html (viewed April 18, 2005). 20. See the Geographic Areas Reference Manual, http://www.census.gov/geo/www/ GARM/Ch13GARM.pdf (viewed June 12, 2007). There are a handful of MSAs or PMSAs for which the constituent counties change between 1990 and 2000 or an MSA was abolished or created. The following tables report results using the MSAs/PMSAs present in each year. We constructed a restricted sample that for the most part held MSA/PMSA boundaries fixed by using only counties that were in the same MSA/PMSA in each of the two years; the estimated levels of and changes in segregation were almost identical. 21. Nonetheless, the results in this paper are generally robust to measuring segregation at the level of the MSA/CMSA metropolitan area rather than the MSA/PMSA level. The only difference is that the increase in black-white segregation is about one-quarter smaller in the first case than in the estimates reported in the following. In addition, we examined our main results for cities disaggregated by quartiles of the population-weighted size distribution, and there was no systematic relationship between city size and changes in segregation along the dimensions we study. 22. These are calculated from Summary File 1 for the 2000 Decennial Census. The population-weighted totals reflect slightly larger MSA/PMSAs. The population weighted 10th percentile is Galveston, TX, with approximately 250,000 residents, and the 90th percentile is Chicago, IL, with approximately 8.3 million residents.

178

Judith Hellerstein, David Neumark, and Melissa McInerney

PMSA the numbers of workers in each category for which we are doing the simulation—for example, blacks and whites—as well as the number of establishments and the size distribution of establishments (in terms of sampled workers). Within a metropolitan area, we then randomly assign workers to establishments, ensuring that we generate the same size distribution of establishments within a metropolitan area as we have in the sample. We do this simulation 100 times and compute the random segregation measures as the means over these 100 simulations. Not surprisingly, the random segregation measures are very precise; in all cases, the standard deviations were trivially small. 5.4 Changes in Segregation With the preceding technical material out of the way, the empirical results can be presented quite concisely. 5.4.1 Segregation by Education The findings for changes in segregation by education are reported in table 5.3. We begin by computing segregation between those with at least some college education and those with at most a high school education. The observed segregation measure for 1990 indicates that, on average, loweducation workers are in workplaces in which 54.2 percent of their coworkers are low education, while high-education workers are in workplaces in which only 34.5 percent are low education, for a difference of 19.7. This is also the effective segregation measure for the national sample because random allocation of workers to establishments anywhere in the country leads to a random coworker segregation measure of zero. When we look within MSAs/PMSAs, randomness generates a fairly small amount of segregation, so the effective segregation measure declines only a little, to 17.3. In the 2000 data, observed segregation is 1.4 percentage points higher (21.1), while random segregation is lower. In combination, then, looking within MSAs/PMSAs, effective segregation by education rises two percentage points, or by 11.3 percent, from 1990 to 2000. In the national data, the increase is smaller, from 19.7 to 21.1 percent, or 7.0 percent.23 The next two panels of table 5.3 report results for two alternative education cutoffs: high school dropouts versus at least a high school degree; and less than a bachelor’s degree versus at least a bachelor’s degree. For the high school dropouts versus at least a high school degree breakdown, the overall national figures indicate an increase in segregation similar to that seen in the first panel of the table; educational segregation increased by 1.7 percentage points (11.1 percent nationally) and by 1.9 percentage points (13.6 percent) within MSAs/PMSAs. When we instead classify workers by whether 23. We remind that reader that when we say “national,” we refer to the MSA/PMSA sample.

Table 5.3

Segregation by education (% low education) 1990 U.S. MSA/PMSA sample (1)

1990 Within MSA/PMSA sample (2)

2000 U.S. MSA/PMSA sample (3)

2000 Within MSA/PMSA sample (4)

Coworker segregation High school degree or less vs. more than high school Observed segregation Low-education workers 54.2 54.2 49.3 High-education workers 34.5 34.5 28.2 Difference 19.7 19.7 21.1 Random segregation Low-education workers 42.9 44.6 35.8 High-education workers 42.9 41.7 35.8 Difference 0 2.9 0 Effective segregation 19.7 17.3 21.1 Percentage point (percent) change, 1990–2000 1.4 (7.0) Less than high school vs. high school degree or more Observed segregation Low-education workers 26.0 26.0 25.5 High-education workers 10.8 10.8 8.6 Difference 15.2 15.2 16.9 Random segregation Low-education workers 12.7 13.8 10.4 High-education workers 12.7 12.6 10.4 Difference 0 1.3 0 Effective segregation 15.2 14.1 16.9 Percentage point (percent) change, 1990–2000 1.7 (11.1)

49.3 28.2 21.1 37.3 35.0 2.3 19.2 2.0 (11.3)

25.5 8.6 16.9 11.3 10.3 1.0 16.0 1.9 (13.6)

Less than bachelor’s degree vs. bachelor’s degree or more Observed segregation Low-education workers 80.7 80.7 77.7 High-education workers 60.6 60.6 54.3 Difference 20.2 20.2 23.4 Random segregation Low-education workers 75.9 76.6 70.8 High-education workers 75.9 73.5 70.8 Difference 0 3.1 0 Effective segregation 20.2 17.6 23.4 Percentage point (percent) change, 1990–2000 3.3 (16.2)

2.8 (16.0)

No. of workers No. of establishments

2,209,908 411,300

1,828,020 317,112

1,828,020 317,112

2,209,908 411,300

77.7 54.3 23.4 71.9 68.2 3.8 20.4

180

Judith Hellerstein, David Neumark, and Melissa McInerney

they have a bachelor’s degree, the increases in segregation are somewhat larger, between 2.8 and 3.3 percentage points, or 16 to 16.2 percent.24 These figures strike us as modest but measurable increases in segregation by education. The direction of change is consistent with the conjecture of Kremer and Maskin (1996), and it is possible that the decade of the 1980s might have experienced even a greater increase in segregation by education, given the sharper increase in schooling-related earnings differentials in that period, although the workforce adjustments may occur relatively slowly. Nonetheless, we may want to be cautious in inferring that the increase in segregation by education is attributable to increased returns to skill. One of the mechanisms for this increase in segregation by education is the decline over the decade in the fraction of workers in the sample with low levels of education—for example, the fraction with at most a high school degree drops from 42.9 percent in 1990 to 35.8 percent in 2000. It is also possible, then, that segregation by skill (rather than measured education) is actually unchanged, but more workers with high unobserved skills have higher education in the 2000 data. 5.4.2 Segregation by Race Evidence on changes in segregation by race is reported in table 5.4. In 1990, the observed segregation measures indicate that blacks, on average, worked with workforces that were 23.7 percent black, whereas the comparable figure for whites was only 5.8 percent, for an observed segregation measure of 17.8. This rose between 1990 and 2000 to 21.8, driven mainly by an increase in the average share black in workplaces where blacks were employed. Nationally, black-white segregation rose 4 percentage points, from 17.8 to 21.8, or an increase of 22.3 percent. Within MSAs/PMSAs, the increase is slightly smaller, at 2.8 percentage points, or 20.3 percent. We interpret these magnitudes as indicating a relatively large increase in workplace segregation by race from 1990 to 2000. 5.4.3 Hispanic-White Segregation Next, table 5.5 reports results for Hispanic-white segregation.25 Observed Hispanic-white segregation is pronounced. In 1990, Hispanic workers, on average, worked in establishments with workforces that were 39.4 percent Hispanic, compared with a 4.5 percent figure for whites. Both of these numbers increased slightly as of 2000, to 40.7 percent and 6 percent, respectively, so that the observed segregation measure remained roughly constant—34.9 percent in 1990 and 34.7 percent in 2000. 24. In Hellerstein and Neumark (forthcoming), we report bootstrapped standard errors for differences in estimates of effective segregation. Differences considerably smaller than the types of increases we find in this paper were strongly significant. 25. Using the 1990 data only, Hellerstein and Neumark (forthcoming) go into considerable detail regarding Hispanic-white segregation, finding that differences in English language skills account for about one-third of this segregation.

Changes in Workplace Segregation in the United States

181

Because of relatively sharp differences in the Hispanic composition of urban areas across the United States, randomness generates a considerable amount of Hispanic-white segregation. This is indicated in the table, where random segregation equals 18.8 in 1990 and 18.0 in 2000. However, again the changes are small so that the change in effective Hispanic-white segregation appears to be relatively minor. Segregation declines in the national Table 5.4

Black-White segregation (% Black)

Coworker segregation Observed segregation Black workers White workers Difference Random segregation Black workers White workers Difference Effective segregation Percentage point (percent) change, 1990–2000 No. of workers No. of establishments

Table 5.5

1990 U.S. MSA/PMSA sample (1)

1990 Within MSA/PMSA sample (2)

2000 U.S. MSA/PMSA sample (3)

2000 Within MSA/PMSA sample (4)

23.7 5.8 17.8

23.7 5.8 17.8

28.7 6.9 21.8

28.7 6.9 21.8

7.1 7.1 0 17.8

11.2 6.8 4.4 14.0

8.8 8.8 0 21.8

14.2 8.3 5.9 16.8

4.0 (22.3)

2.8 (20.3)

1,893,034 360,072

1,893,034 360,072

1,618,876 285,988

1,618,876 285,988

Hispanic-White segregation (% Hispanic)

Coworker segregation Observed segregation Hispanic workers White workers Difference Random segregation Hispanic workers White workers Difference Effective segregation Percentage point (percent) change, 1990–2000 No. of workers No. of establishments

1990 U.S. MSA/PMSA sample (1)

1990 Within MSA/PMSA sample (2)

2000 U.S. MSA/PMSA sample (3)

2000 Within MSA/PMSA sample (4)

39.4 4.5 34.9

39.4 4.5 34.9

40.7 6 34.7

40.7 6 34.7

6.9 6.9 0 34.9

24.4 5.6 18.8 19.8

9.2 9.2 0 34.7

25.5 7.5 18.0 20.4

–0.2 (–0.4)

0.6 (3.0)

1,906,878 373,006

1,906,878 373,006

1,625,953 293,989

1,625,953 293,989

182

Judith Hellerstein, David Neumark, and Melissa McInerney

data by 0.2 percentage point, or by less than 1 percent. And within urban areas, segregation increases slightly, from 19.8 to 20.4, or by only 3 percent. Overall, then, both the small magnitudes and the differences in results across and within urban areas lead us to conclude that little changed with respect to Hispanic-white workplace segregation between 1990 and 2000. 5.4.4 Sex Segregation Finally, we turn to segregation by sex. A priori, we might expect to find substantial declines in this form of segregation because of the declining differences in the types of jobs done by men and women (Wells 1998). As table 5.6 reports, in 1990 women, on average, worked in establishments with workforces that were 59.9 percent female, as compared with establishments in which men worked, which were 36.2 percent female. Thus, observed segregation was 23.6. As of 2000, the increase in the share female with which men work increased relatively sharply, from 36.2 to 40.2, and as a result observed segregation fell to 20.4. Random segregation by sex is relatively trivial because neither men nor women constitute a very small share of the workforce. As a result, the change in effective segregation is close to the change in observed segregation. In particular, effective segregation by sex declined from 23.6 to 20.4, or 13.7 percent, on a national basis. And virtually the same decline, 3.2 percentage points or 13.6 percent, is estimated within urban areas because, of course, the distributions of men and women across cities are similar. We view the magnitude of these changes in sex segregation as suggesting a substantive decline over the decade. One possible explanation for the overall decline in sex segregation is convergence in the occupational distributions of men and women, rather than a reduction in segregation across workplaces even for men and women in the same occupation. To address this possibility, following the methods in Hellerstein and Neumark (forthcoming), we construct “conditional” random segregation measures, where we simulate segregation holding the distribution of workers by occupation fixed across workplaces. So, for example, if an establishment in our sample is observed to have three workers in occupation A, then three workers in occupation A will be randomly allocated to that establishment. As before, we compute the average (across the simulations) simulated fraction of coworkers who are female for females, denoting this FFC, and the average (across the simulations) simulated fraction of coworkers who are female for males, denoting this MFC. The difference between these two is denoted CWC, and we define the extent of “effective conditional segregation” to be CWO – CWC  100, 100 – CWR where CWR is the measure of random segregation obtained when not conditioning on occupation. A conditional effective segregation measure of

No. of workers No. of establishments

1,828,020 317,112

47.7 47.2 0.5 23.3

47.4 47.4 0 23.6

1,828,020 317,112

59.9 36.2 23.6

2,209,908 411,300

–3.2 (–13.6)

–3.2 (–13.7)

2,209,908 411,300

50.7 50.3 0.4 20.1

60.6 40.2 20.4

2000 Within MSA/PMSA sample (4)

50.5 50.5 0 20.4

60.6 40.2 20.4

2000 U.S. MSA/PMSA sample (3)

Unconditional 1990 Within MSA/PMSA sample (2)

59.9 36.2 23.6

1990 U.S. MSA/PMSA sample (1)

Segregation by sex (% female)

Coworker segregation Observed segregation Female workers Male workers Difference Random segregation Female workers Male workers Difference Effective segregation Percentage point (percent) change, 1990–2000 Fraction of sex segregation accounted for by occupation

Table 5.6

1,828,020 317,112

55.4

54.4 41.1 13.3 10.4

59.9 36.2 23.6

1990 Within MSA/PMSA sample (5)

2,209,908 411,300

61.2

–2.6 (–24.8)

56.8 44.1 12.6 7.8

60.6 40.2 20.4

2000 Within MSA/PMSA sample (6)

Conditional on 3-digit occupation

184

Judith Hellerstein, David Neumark, and Melissa McInerney

zero would imply that all of the effective segregation between women and men can be attributed to differences in the occupations employed by various establishments (“occupational segregation”), coupled with differences in the occupational distributions of women and men. Conversely, a conditional effective segregation measure equal to that of the (unconditional) effective segregation measure would imply that none of the effective segregation between women and men can be attributed to occupational segregation across workplaces. Columns (5) and (6) of table 5.6 report the results of doing this calculation based on a consistent occupation classification across 1990 and 2000, as developed in Meyer and Osborne (2005), which is approximately at the three-digit level.26 We do this only for the within MSA/PMSA sample because central to this analysis is the ability to randomly distribute workers to different establishments, and it makes more sense to do this within the urban areas in which workers commute. The estimates for 1990, in column (5), indicate that a substantial fraction (nearly 50 percent) of the effective segregation of women from men is attributable to differences in the occupational distribution; conditional on occupation, effective segregation by sex falls from 23.3 (column [2]) to 10.4. In the 2000 data, reported in column (6), the effect of occupation is a little bit more pronounced, accounting for 61.2 percent of effective segregation. Finally, conditional on occupation, sex segregation within MSAs/PMSAs declines over time by 2.6 percentage points (from 10.4 to 7.8); in absolute terms, this is similar to the decline in unconditional segregation, but because effective segregation conditional on occupation (in 1990) was only about 45 percent as large as the unconditional effective segregation measure, the decline in conditional segregation between columns (5) and (6) represents a much larger percentage decline—24.8 percent. Altogether, these results suggest that the decline in sex segregation over the decade is not being driven by the increased propensity of women to work in the same occupations as men. 5.5 The Impact of Changing Establishment and Industry Composition Changes in segregation can arise due to a multitude of factors, some of them compositional, such as the changing occupational distribution of women as discussed in the previous section. In this section, we explore the robustness of our full-sample results to two other types of potentially important compositional changes. First, we explore whether the changes in segregation are due to the changing composition of establishments by recalculating our segregation indexes for only the sample of establishments 26. There are nontrivial differences in occupation codes at the three-digit level between 1990 and 2000. The structure of occupation codes at the one-digit level changed even more dramatically between 1990 and 2000, so we do not attempt a concordance at this higher level of aggregation.

Changes in Workplace Segregation in the United States

185

that exist in both the 1990 and 2000 Restricted DEED samples (corresponding to columns [3] and [6] of table 5.1).27 Ideally, we would like to isolate the separate roles of establishment entry and exit—that is, births of new establishments and deaths of existing ones. However, given that we only match some establishments, we cannot necessarily distinguish births and deaths from matches and nonmatches. But assuming that matching is random with respect to segregation, focusing on the set of establishments that are in both samples is informative about the combined roles of establishment entry and exit. Second, we explore the robustness of our changes in segregation to changes in the industry mix of employment over the decade by reweighting the segregation indexes for 2000 to reflect the industrial composition of employment at the one-digit level that exists in our 1990 data. This is a little more complicated. First, because we are interested in calculating withinMSA indexes, it is actually the within-MSA industry composition that we need to hold fixed at 1990 levels. As a result, we include in the sample only MSAs that exist in both years. Second, we exclude mining because mining makes up such a trivial proportion of employment that there are some MSAs that have matched workers in mining in 1990 but not in 2000. To understand how we construct changes in segregation over the decade while holding the distribution of employment across industries within MSAs fixed at 1990 levels, consider again the example of ethnic segregation we discussed in section 5.3. Obviously, we compute HH (the isolation index) and WH (the exposure index) for 1990 in the same way we did previously because no adjustment needs to be made when accounting for the 1990 industry composition. In order to compute HH for 2000 with industry composition fixed as of 1990, we compute the isolation index separately for each industry/MSA pair in 2000.28 We then take a weighted average across industries of these isolation indexes, where the weight is the product of two components: the fraction of total Hispanic employment (in this example) that works in that industry/MSA pair in 2000, and the ratio of the employment share in the industry/MSA pair in 1990 relative to 2000. The fraction of Hispanic employment serves to aggregate up the industry/MSA-specific isolation indexes to the full-sample isolation index (and, if used alone to weight up the industry/MSA-specific indexes would yield the 2000 unadjusted isolation index), while the ratio of the employment shares adjusts the data appropriately to reflect the composition of employment in 1990 across industries. For the exposure index, WH, we do the same thing, calculating a 27. By restricting the sample to establishments that exist in the Restricted DEED samples in both years, we drop some very small MSAs from some of the samples we used to calculate segregation indexes in earlier tables, in cases where there are no matched workers for whom to calculate indexes across the two years. 28. For the random segregation indexes, the industry used is the random industry to which the worker is assigned.

186

Judith Hellerstein, David Neumark, and Melissa McInerney

separate exposure index for each industry/MSA pair and then weighting by the product of the industry employment share ratio times the fraction of white employment in that pair in 2000. Because the fraction Hispanic in an industry MSA/pair may differ from the fraction white in that same industry/MSA pair, the reweighting may have differential effects on the exposure and isolation indexes. As a consequence, adjusting for industry employment changes over the decade will have the largest impact on measured changes in segregation when there has been differential employment growth in industries with a large share Hispanic coupled with a large difference between the share of Hispanic and the share of white employment in the industry (or if there is a large difference between the isolation and exposure indexes).29 The results of these alternative computations are presented in condensed form in table 5.7, where we report only the within-MSA effective segregation measures in each year and the changes over the decade. In the first panel of table 5.7, we report results for coworker segregation by high school degree status. In column (1), we first report the within-MSA effective segregation measure in 1990 of 17.3 (from table 5.3). Following that number, we report the corresponding figure for the sample of establishments that existed both in 1990 and 2000, finding that coworker segregation by high school degree status in 1990 is somewhat lower, at 15.7. The fixed-industrycomposition coworker segregation measure for 1990 is 17.3, identical to that for the full sample.30 In column (2), we report the coworker segregation measures for 2000. For the fixed-establishment sample, coworker segregation by high school degree status is 17.0, 2.2 percentage points lower than for the full sample, and for the results holding industry composition fixed, the coworker measure is slightly higher, at 20.3. Overall, the change over the decade of 2 percentage points for the full sample is close to the 1.4 percentage point increase for the fixed-establishment sample, and the increase holding industry composition fixed is a bit larger, at 3.1 percentage points. In general, though, the observed increase in coworker segregation for the full sample over the decade is robust to the changing mix of establishments and industries. In the second and third panels of table 5.7, we report the results for the alternative education cutoffs. The results again reflect some small differences across the sample of establishments and mix of industries, and the overall qualitative results again point to increases in segregation by education over the decade. 29. This turns out to be quite significant in our calculations for changes in sex segregation holding the industry composition of employment fixed, where the services industry grew rapidly and is also heavily female. 30. Because we exclude workers in mining and workers in MSAs that were not defined as such in 1990 and 2000, the results for 1990 can be slightly different than we report in the full sample in table 5.3.

Changes in Workplace Segregation in the United States Table 5.7

187

Alternative coworker segregation calculations

Segregation by education High school degree or less vs. more than high school Full sample, table 5.3 Establishments present in 1990 and 2000 Fixed industry composition Less than high school vs. high school degree or more Full sample, table 5.3 Establishments present in 1990 and 2000 Fixed industry composition Less than bachelor’s degree vs. bachelor’s degree or more Full sample, table 5.3 Establishments present in 1990 and 2000 Fixed industry composition Black-White segregation Full sample, table 5.4 Establishments present in 1990 and 2000 Fixed industry composition Hispanic-White segregation Full sample, table 5.5 Establishments present in 1990 and 2000 Fixed industry composition Segregation by Sex Unconditional Full sample, table 5.6 Establishments present in 1990 and 2000 Fixed industry composition

1990 Within MSA/PMSA sample, effective segregation (1)

2000 Within MSA/PMSA sample, effective segregation (2)

Percentage point (percent) change, 1990–2000 (3)

17.3

19.2

2.0 (11.3)

15.7 17.3

17.0 20.3

1.4 (8.9) 3.1 (17.8)

14.1

16.0

1.9 (13.6)

11.4 13.8

12.7 15.8

1.2 (10.7) 2.0 (14.3)

17.6

20.4

2.8 (16.0)

15.4 17.6

17.4 21.8

2.0 (12.8) 4.2 (24.0)

14.0

16.8

2.8 (20.3)

11.2 14.1

12.6 14.7

1.4 (12.7) 0.6 (4.6)

19.8

20.4

0.6 (3.0)

16.5 19.1

15.6 22.0

–0.9 (–5.6) 2.9 (15.3)

23.3

20.1

–3.2 (–13.6)

25.2 23.4

23.0 14.4

–2.3 (–8.9) –9.0 (–38.3)

Note: Mining is excluded for “Full sample” and “fixed industry composition.”

Racial segregation increased over the decade for the full sample by 2.8 percentage points (20.3 percent), but increased by only about half that much for the sample of establishments that exist in both years. This means that new establishments in 2000 are characterized by more racial segregation than establishments that existed in 1990. Moreover, holding the industry composition of employment fixed at 1990 levels, racial segregation

188

Judith Hellerstein, David Neumark, and Melissa McInerney

increased by a much smaller amount over the decade—0.6 percentage points (4.6 percent). The fact that newer establishments and the industries that are gaining in employment over the decade are also more segregated by race in 2000 than older establishments and declining industries could portend continuing increases in racial segregation. Interestingly, the results are somewhat different for Hispanic-white segregation, as we report in the fifth panel of table 5.7. In the overall sample, coworker segregation increased relatively little over the decade, by only 0.6 percentage points (3 percent). For the sample of establishments that exist in both years, coworker segregation actually fell a little, from 16.5 to 15.6, whereas for the fixed-industry-composition, the coworker segregation measure rose over the decade from 19.1 to 22.0. While none of these results point to major differences, it appears that the changing industry mix served to decrease Hispanic-white segregation, while the entry and exit of establishments seems to have worked in the opposite direction. As reported in the last panel of table 5.7, the results are most notably different for sex segregation—in particular, with respect to the role of industry composition. First, for the full sample, coworker segregation fell by 3.2 percentage points over the decade, whereas for the sample of continuing establishments, it fell by 2.3 percentage points. Because the baseline coworker segregation measure in 1990 for the continuing establishments sample is slightly higher (25.2) than for the full sample (23.3), on a percentage basis segregation actually declined somewhat more for the full sample, but the difference is small. However, a much sharper difference arises when comparing the change over the decade to that obtained holding the distribution of employment across industries fixed at 1990 levels. In particular, coworker segregation in 2000 is 20.1 in the full sample, but only 14.4 in the fixed-industry-composition results. As a result, coworker segregation for the fixed-industry-composition calculation falls over the decade by a full 9 percentage points, or 38.3 percent. Industry composition has such a strong influence on changes in measured sex segregation because there was very sharp employment growth in services, which is a highly sex segregated industry with a high share of female employment. Thus, absent the growth in services (which is what we mimic by holding the industry composition of employment fixed), sex segregation would have declined by considerably more. To see this, table 5.8 presents detailed information on isolation and exposure indexes and observed segregation by industry, as well as the distribution of employment of men and women across industries, and industry employment growth over the decade. The table shows, first, that with the exception of transportation, which is a relatively small industry, services is the most segregated industry in both years. The percentage point decline in observed segregation is relatively similar across industries, with the exception of wholesale. In addition, the services industry was the largest employer of women in both years, accounting for 35 percent of female employment in

Changes in Workplace Segregation in the United States Table 5.8

Observed sex segregation, by industry, within MSA/PMSA

Observed segregation

Construction Female Male Difference Manufacturing Female Male Difference Transportation Female Male Difference Wholesale Female Male Difference Retail Female Male Difference FIRE Female Male Difference Services Female Male Difference

189

1990 (1)

2000 (2)

23.7 21.1 2.5

21.6 23.9 –2.3

43.7 28.2 15.5

41.2 29.8 11.4

49.0 29.4 19.6

48.8 32.2 16.7

40.2 31.9 8.3

38.6 31.6 6.9

59.3 42.9 16.4

57.1 44.7 12.4

69.3 61.3 8.0

69.1 63.7 5.4

74.2 55.1 19.2

72.7 58.4 14.3

Percentage point (percent) change, 1990–2000 (3)

Share of female or male employment in industry

Ratio of 1990 to 2000 industry employment (6)

1990 (4)

2000 (5)

0.02 0.05

0.02 0.06

0.92

0.24 0.44

0.17 0.35

1.29

0.04 0.07

0.04 0.07

1.05

0.06 0.10

0.03 0.07

1.50

0.17 0.14

0.20 0.20

0.80

0.12 0.05

0.09 0.05

1.24

0.35 0.15

0.44 0.21

0.76

–4.9 (–191.4)

–4.1 (–26.3)

–3.0 (–15.1)

–1.4 (–16.3)

–4.0 (–24.3)

–2.6 (–32.7)

–4.8 (–25.1)

1990 and 44 percent of female employment in 2000. So, for example, services alone accounts for half of the isolation index in 2000.31 Services was also the second largest employer of men in both years, but well behind manufacturing. Employment of both men and women in services grew sharply over the decade. This is reflected in the distribution of men and women across industries by year (columns [4] and [5]), as well as in the ratio of overall employment in 1990 relative to 2000, as reported in column (6); services has the lowest ratio (0.76), corresponding to the sharpest growth. 31. This can be seen by multiplying the isolation index for services of 72.7 percent by the employment share of services in female employment of 44 percent, as reported in table 5.8, accounting for 53 percent of the overall isolation index in 2000 of 60.7 corresponding to the sample used in that table.

190 Table 5.9

Judith Hellerstein, David Neumark, and Melissa McInerney Sex segregation including and excluding services, within MSA/PMSA

Including services Observed segregation Female Male Difference Random segregation Female Male Difference Effective segregation Percentage point (percent) change, 1990–2000 No. of workers No. of establishments Excluding services Observed segregation Female Male Difference Random segregation Female Male Difference Effective segregation Percentage point (percent) change, 1990–2000 No. of workers No. of establishments

1990

2000

2000, fixed industry composition

59.9 36.2 23.7

60.7 40.2 20.5

55.1 40.3 14.7

47.7 47.2 0.5 23.4

50.7 50.4 0.4 20.2

50.7 50.4 0.4 14.4

–3.2 (–13.5)

–9.0 (–38.3)

1,739,063 301,029

2,151,566 398,958

2,151,566 398,958

52.1 32.9 19.2

51.2 35.4 15.9

48.9 35.1 13.8

41.0 40.5 0.5 18.8

42.2 41.9 0.3 15.7

42.2 41.9 0.3 13.6

–3.1 (–16.5)

–5.2 (–27.7)

1,450,311 289,206

1,450,311 289,206

1,310,125 236,412

Note: Mining is excluded.

As a result of the fact that services is a relatively highly sex segregated industry, coupled with the fact that it is a heavily female industry that grew tremendously over the decade, the services industry plays a large role in overall changes in sex segregation. To reinforce the importance of the growth in services employment in mitigating the decline in sex segregation, table 5.9 shows our calculations of effective sex segregation and how it changed over the decade, with and without holding the industry composition of employment fixed and with and without including services. The top panel shows calculations for all industries, echoing the earlier results.32 In 32. As noted earlier, the estimates allowing the industry composition to change are slightly different from in table 5.6 because of changes in the sample to do the calculation holding industry composition fixed.

Changes in Workplace Segregation in the United States

191

the bottom panel, however, services is simply dropped from the calculation, and the difference in the change in segregation from holding industry composition fixed is only about half as large.33 5.6 Robustness to Alternative Segregation Measures Finally, table 5.10 compares the estimates for the key results using the coworker segregation measure and the Duncan index to see how robust the results are to alternative segregation measures. For segregation by education, the measure of effective segregation in each year and for each educational split is quite similar. The changes in segregation are also similar, with the only exception that the Duncan index points to a larger increase in segregation between those with less than a high school degree versus those with a high school degree or more. The results for black-white and Hispanic-white segregation all show increases over the decade, although the increase in the Duncan index relative to the coworker measure is a little smaller for blackwhite segregation and a little larger for Hispanic-white segregation. The estimates for sex segregation corresponding to table 5.6 (i.e., unconditional, and conditional on occupation), are very similar for the two measures, with both showing marked declines, and of similar magnitudes. However, holding industry composition fixed appears to have much less impact using the Duncan index. Given that the importance of industry composition using the coworker segregation measure is derived in part from the large share female in services coupled with the strong growth of that industry, it is not surprising that the effect of holding industry constant has less of an impact when using the Duncan index because this index is not sensitive to simple changes in a group’s representation in the workforce that are distributed across establishments in proportion to their original distribution. We argued earlier that we have some preference for the coworker measure over indexes like the Duncan index, most importantly because we think the variation in the coworker measure in response to simple changes in the share of a particular group in the workforce represents meaningful changes in workplace segregation, but also for the more technical reasons discussed in section 5.3. In general, though, the conclusions that can be drawn from the two segregation measures are qualitatively similar and, in particular, the directions of the changes across the decade are always the same. Given the differing properties of the two measures, however, the quantitative answers obviously differ somewhat. Nonetheless, as a summary measure of the comparability of the estimates, the last row of the table shows that the estimated percentage point and percent changes are highly correlated across the two indexes (0.78 and 0.83, respectively), computed across all of the estimates reported in the tables. 33. Note that there was also strong growth in retail, another industry that is relatively sexsegregated.

Table 5.10

Comparisons of results for coworker segregation and Duncan Indexes, effective segregation, within MSA/PMSA Effective segregation 1990 (1)

2000 (2)

Percentage point change, 1990–2000 (3)

17.3 25.3

19.2 28.6

2.0 3.3

11.3 13.0

14.1 29.8

16.0 37.5

1.9 7.7

13.6 25.7

17.6 26.3

20.4 28.5

2.8 2.2

16.0 8.5

Table 5.4 Black-white Coworker Duncan

14.0 18.4

16.8 20.9

2.8 2.5

20.3 13.5

Table 5.5 Hispanic-white Coworker Duncan

19.8 19.8

20.4 23.2

0.6 3.5

3.0 17.5

23.3 31.6

20.1 28.4

–3.2 –3.2

–13.6 –10.2

10.4 11.3

7.8 8.5

–2.6 –2.8

–24.8 –24.7

23.4 31.7

14.4 28.5

–9.0 –3.2 .78

–38.3 –11.3 .83

Table 5.3 High school degree or less vs. more than high school Coworker Duncan Less than high school vs. high school degree or more Coworker Duncan Less than bachelor’s degree vs. bachelor’s degree or more Coworker Duncan

Table 5.6 Male-female Unconditional Coworker Duncan Conditional on 3-digit occupation Coworker Duncan Table 5.7 Male-female Unconditional, fixed industry composition Coworker Duncan Correlation between indexes Note: See notes to corresponding tables.

Percent change, 1990–2000 (4)

Changes in Workplace Segregation in the United States

193

5.7 Conclusions We present evidence on changes in workplace segregation by education, race, ethnicity, and sex. For this analysis, we use the newly-constructed 2000 Decennial Employer-Employee Dataset (DEED). The 2000 DEED, like the 1990 DEED, provides new opportunities to study workplace segregation at the establishment level. More significantly, by pairing the two we are able to present what we believe are the first estimates of changes in workplace segregation based on 2000 Census data. These estimates provide evidence that is complementary to that on changes in residential segregation in the decade between the Censuses. Moreover, we believe that evidence on workplace segregation and how it has changed is likely to be more informative about social interactions between groups (with reference to race, ethnicity, and sex), and directly informative about hypotheses regarding changes in workplace segregation by skill. The evidence indicates that racial and ethnic segregation at the workplace level remains quite pervasive. For example, if we compare black and white workers, the difference in the share black among the workforce at the establishments where they work is around 22 percentage points. If we compare Hispanics and whites, the difference is about 50 percent larger. At the same time, there is fairly substantial segregation by skill, as measured by education. In other work (Hellerstein and Neumark [forthcoming], using only the 1990 DEED), we explore the extent to which racial and ethnic segregation is attributable to skill differences between blacks and whites or Hispanics and whites; in the latter case, we focus on language skills. Only for the latter is there evidence that skill differences play a substantial role, explaining about one-third of Hispanic-white segregation. More significantly, putting together the 1990 and 2000 data, we find no evidence of declines in workplace segregation by race and ethnicity. Hispanic-white segregation was largely unchanged, while black-white segregation increased by about 3 to 4 percentage points, or about 20 percent. This increase in racial segregation was reinforced by the entry and exit of establishments and by the changing industry composition of employment, suggesting that there may be forces at work that will lead to the persistence of or even increases in racial segregation. Over this decade, segregation by education also increased, by about 2 to 3 percentage points, or 11 to 16 percent. This increase is consistent with conjectures that rising returns to skill might generate more segregation by skill, although it could also be attributable to rising education levels among workers with more unobserved skills but with an unchanging pattern of segregation based on these skills.34 34. Coupled with the earlier findings suggesting that black-white segregation is largely unrelated to education differences, this likely has little if anything to do with the increase in workplace segregation by race.

194

Judith Hellerstein, David Neumark, and Melissa McInerney

To the extent that declines in segregation are positive developments, the one bright spot is the decline in workplace segregation by sex, which fell about 3 percentage points, or 14 percent, from 1990 to 2000. Changes in the occupational distribution of men and women did not play a major role in this decline in segregation. If we hold the distribution of men and women across three-digit occupations fixed, the absolute decline in segregation is roughly the same, although it is larger in percentage terms. On the other hand, shifts in the industry composition of employment worked against the decline in sex segregation, as the fastest-growing industry was services, which is also one of the most sex-segregated and most heavily female industries. What this implies for future changes in sex segregation depends on whether the shifts in industrial composition continue as in the recent past or change course, on changes in the distribution of women across industries, and on changes in sex segregation across establishments within industries.

References Autor, David H., Lawrence F. Katz, and Melissa S. Kearney. 2005. Trends in U.S. wage inequality: Re-assessing the revisionists. NBER Working Paper no. 11627. Cambridge, MA: National Bureau of Economic Research. Bayard, Kimberly, Judith Hellerstein, David Neumark, and Kenneth Troske. 1999. Why are racial and ethnic wage gaps larger for men than for women? Exploring the role of segregation using the New Worker-Establishment Characteristics Database. In The creation and analysis of employer-employee matched data, ed. J. Haltiwanger, J. Lane, J. R. Spletzer, J. Theeuwes, and K. Troske, 175–203. Amsterdam: Elsevier Science. ———. 2003. New evidence on sex segregation and sex differences in wages from matched employee-employer data. Journal of Labor Economics 21 (4): 887–922. Blau, Francine D. 1977. Equal pay in the office Lexington, MA: Heath. Boisso, Dale, Kathy Hayes, Joseph Hirschberg, and Jacques Silber. 1994. Occupational segregation in the multidimensional case. Journal of Econometrics 61 (1): 161–71. Carrington, William J., and Kenneth Troske. 1997. On measuring segregation in samples with small units. Journal of Business and Economic Statistics 15 (4): 402–9. ———. 1998a. Interfirm racial segregation and the black/white wage gap. Journal of Labor Economics 16 (2): 231–60. ———. 1998b. Sex segregation in U.S. manufacturing. Industrial and Labor Relations Review 51:445–64. Cortese, Charles F., R. Frank Falk, and Jack K. Cohen. 1976. Further considerations on the methodological analysis of segregation indices. American Sociological Review 51 (4): 630–7. Duncan, Otis D., and Beverly Duncan. 1955. A methodological analysis of segregation indices. American Sociological Review 20 (2): 210–7. Echenique, Federico, and Roland Fryer. 2005. On the measurement of segregation. NBER Working Paper no. 11258. Cambridge, MA: National Bureau of Economic Research.

Changes in Workplace Segregation in the United States

195

Estlund, Cynthia. 2003. Working together: How workplace bonds strengthen a diverse democracy. New York: Oxford University Press. Foster, Lucia, John Haltiwanger, and C. J. Krizan. 1998. Aggregate productivity growth: Lessons from microeconomic evidence. NBER Working Paper no. 6803. Cambridge, MA: National Bureau of Economic Research. Glaeser, Edward L., and Jacob L. Vigdor. 2001. Racial segregation in the 2000 Census: Promising news. The Brookings Institution Survey Series. Washington, DC: Brookings Institution, April. Groshen, Erica L. 1991. The structure of the female/male wage differential: Is it who you are, what you do, or where you work? Journal of Human Resources 26 (3): 457–72. Hellerstein, Judith, and David Neumark. 2003. Ethnicity, language, and workplace segregation: Evidence from a new matched employer-employee data set. Annales d’Economie et de Statistique 71–72:19–78. ———. 2007. Production function and wage equation estimation with heterogeneous labor: Evidence from a new matched employer-employee data set. In Hard to Measure Goods and Services: Essays in Honor of Zvi Griliches, eds. E. R. Berndt and C. R. Holten, 31–71. Chicago: University of Chicago Press. ———. Forthcoming. Workplace segregation in the United States: Race, ethnicity, and skill. Review of Economics and Statistics. Higgs, Robert. 1977. Firm-specific evidence on racial wage differentials and workforce segregation. American Economic Review 67 (2): 236–45. Iceland, John, and Daniel H. Weinberg. 2002. Racial and ethnic segregation in the United States: 1980–2000. U.S. Census Bureau, Census 2000 Special Reports. http://www.census.gov/hhes/www/housing/housing_patterns/pdf/censr-3.pdf. James, Daniel R., and Karl E. Taeuber. 1985. Measures of segregation. In Sociological methodology, ed. N. B. Tuma, 1–32. San Francisco: Jossey-Bass. Jarmin, Ron S., and Javier Miranda. 2002. The longitudinal business database. CES Working Paper no. CES-WP-02-17. Washington, DC: Center for Economic Studies. King, Mary C. 1992. Occupational segregation by race and sex, 1940–1988. Monthly Labor Review 115 (4): 30–7. Kremer, Michael, and Eric Maskin. 1996. Wage inequality and segregation by skill. NBER Working Paper no. 5718. Cambridge, MA: National Bureau of Economic Research. McConville, Shannon, and Paul Ong. 2001. Examining residential segregation patterns. The Ralph and Goldy Lewis Center for Regional Policy Studies, Discussion Paper, The University of California at Los Angeles. Meyer, Peter B., and Anastasiya M. Osborne. 2005. Proposed category system for 1960–2000 Census occupations. BLS Working Paper no. 383. Washington, DC: Bureau of Labor Statistics, September. Watts, Martin J. 1995. Trends in occupational segregation by race and gender in the U.S.A., 1983–92: A multidimensional approach. Review of Radical Political Economics 27 (4): 1–36. Wells, Thomas. 1998. Change in occupational sex segregation during the 1980s and 1990s. Center for Demography and Ecology. Working Paper no. 98-14. Madison: University of Wisconsin. Winship, Christopher. 1977. A revaluation of indexes of residential segregation. Social Forces 55 (4): 1058–66.

6 The Effect of Tuition Reimbursement on Turnover A Case Study Analysis Colleen Flaherty Manchester

6.1 Introduction Becker’s (1964) seminal work on investment in human capital makes a fundamental distinction between general and firm-specific skills, which has implications for investment and employee turnover. Firm-specific human capital is defined as having value only to the current employment relationship, while general human capital is valuable to both current and potential employers. Becker’s theory predicts that employees will bear the full cost of general skills training—either by paying for training directly or by accepting lower wages during training periods—because employers face the threat of not capturing the return on their investment due to “poaching” of trained employees by other employers. In a competitive labor market, workers have the incentive to invest efficiently in general human capital because they receive a wage equal to the value of their marginal Colleen Flaherty Manchester is an assistant professor at the Carlson School of Management, University of Minnesota. I would like to thank the participants at the Conference for the Analysis of Firms and Employees (CAFE) for their comments and suggestions as well as the editors and reviewers of this conference volume. In addition to the conference sponsors, this research was made possible by the Hawley-Shoven Fellowship Fund, through a grant to the Stanford Institute for Economic Policy Research. I would like to thank my advisors, John Shoven, Michael Boskin, John Pencável, and Edward Lazear, as well as members of the labor reading group at Stanford University for their comments and feedback. I would also like to thank Harley Frazis, Merrisa Piazza, Maury Gittleman, and James Spletzer at the U.S. Bureau of Labor Statistics as well as the Human Resource Department at the case study institution for access to the data used in this analysis. The views presented in this paper are mine alone and do not reflect those of my advisors, the aforementioned organizations, or the National Bureau of Economic Research.

197

198

Colleen Flaherty Manchester

product. In the case of investment in firm-specific human capital, employers and employees share the costs. Neither party is willing to bear the full amount due to the risk of opportunistic behavior by the other. The employer and the employee share the surplus, or rents, from the investment; the relative bargaining power of the two parties determines how these rents are allocated. This standard theory on investment in human capital has implications for turnover. Investment in firm-specific human capital reduces turnover because rents accrue only if the employment relationship is maintained. However, this result does not hold for investments in general human capital because these skills are transferable across employers. According to standard theory, offering employees general skills training would increase turnover. Despite the predictions of this theory, recent empirical studies show that firms provide general training to their workers and often argue that firms bear part of the cost.1 Tuition reimbursement programs are an example of general skills training provided by firms. Employers reimburse employees for direct costs of coursework taken at accredited academic institutions. Because instruction and degree accreditation occur at third-party institutions, skills acquired are transferable—as well as observable—to many potential employers. Hence, tuition reimbursement programs closely resemble general skills training as described by Becker (1964). A primary reason firms give for offering these programs is to reduce turnover, which challenges standard human capital theory.2 This chapter examines empirically whether employees who participate in tuition reimbursement have higher retention rates than nonparticipants using a case study analysis of workers at a nonprofit institution. Results from the case study indicate that participation reduces employee turnover. Hence, the firm’s motivation for offering this program is supported by this analysis: general skills training increases retention. The remainder of this chapter is organized as follows. Section 6.2 documents the prevalence of tuition reimbursement programs using the Survey of Employer-Provided Training, 1995 (SEPT95) and describes the typical characteristics of these programs. Section 6.3 reviews previous studies of tuition reimbursement programs, while the case study analysis is presented in section 6.4. Section 6.5 discusses the results from the case study, and section 6.6 concludes the chapter.

1. These include, but are not limited to, Loewenstein and Spletzer (1999a, b), Acemoglu and Pischke (1998, 1999a, b), Autor (2001), and Cappelli (2004). 2. Increased retention is a response given by human resource professionals in interviews with the author. Cappelli (2004) and the Corporate Leadership Council (2003) report the same finding.

The Effect of Tuition Reimbursement on Turnover

199

6.2 Background on Tuition Reimbursement Programs 6.2.1 Program Prevalence Employer-provided tuition reimbursement programs are widespread and constitute a nontrivial part of nonwage compensation. One of the few data sets that collects information on tuition reimbursement programs is SEPT95. This survey collects information on employer-provided training practices for a cross section of establishments. The survey was conducted by the Bureau of Labor Statistics (BLS), part of the U.S. Department of Labor, from May to August of 1995, with the purpose of collecting nationally representative data (Bureau of Labor Statistics 1996). The sample represents private, nonagriculture establishments employing fifty or more workers. The results from SEPT95 show that a substantial fraction of firms offer tuition reimbursement: 61 percent of establishments employing fifty or more workers offer tuition reimbursement programs. Using the 1994 National Employer Survey of Educational Quality in the Workforce (NES-EQW), Black and Lynch (1998) report that 47 percent of firms employing twenty or more employees offer tuition reimbursement programs. The estimates from SEPT95 and 1994 NES-EQW are comparable because larger establishments are more likely to offer a tuition reimbursement program (Frazis et al. 1998; Frazis, Gittleman, and Joyce 2000). Examining the access of workers to this program, over three-quarters of the employees who work in establishments represented in SEPT95 are offered tuition reimbursement by their employer. In addition, the amount spent on these programs is substantial: establishments represented in SEPT95 spent $2.8 billion in 1994 on tuition reimbursement.3 The trade magazine Workforce Management estimates that companies paid $10 billion toward tuition reimbursements in 2003.4 Hence, expenditures on tuition reimbursement programs represent a significant source of investment in general skills of employees and appear to be on the rise. However, these programs are relatively unexamined in the academic literature. This chapter examines the effect of participation in the tuition reimbursement program offered by the case study institution on the propensity of an employee to separate from his or her employer. 6.2.2 Program Characteristics Tuition reimbursement programs typically consist of three components: (1) a maximum reimbursement amount; (2) an eligibility requirement; and (3) a reimbursement policy based on academic performance. The first 3. The confidence interval for this estimate ranges from $2.6 billion to $3.0 billion (1994 U.S. dollars). 4. See Workforce Management, May 1, 2004.

200

Colleen Flaherty Manchester

characteristic is affected by the tax-advantage status of these programs. Reimbursements from employer-provided programs are exempt from income taxation under Section 127 of the U.S. Internal Revenue Code.5 The maximum income exclusion for a single calendar year is $5,250, and reimbursements under this level are exempt from federal income tax, payroll tax, and state income tax. A 2002 survey by Eduventures of human resource professionals and managers at over 500 firms finds that 70 percent of firms offering a tuition reimbursement program cap annual reimbursement, and over half of these firms (57 percent) have maximums that exceed $4,000 (Newman and Stein 2003).6 Table 6.1 shows the distribution of reimbursement maximums from the Eduventures’ survey. The majority of firms choose maximums below or equal to the maximum annual tax exclusion, $5,250, but a sizable fraction of firms have reimbursement maximums that exceed the tax exempt limit or have no maximum reimbursement amount. Among firms offering tuition benefits, nearly 40 percent offer reimbursements beyond the level that receives tax-advantaged status (i.e., amounts greater than $5,250). The fact that firms set annual limits in excess of the tax exempt amount provides evidence that these programs are not solely offered because of their tax-advantaged status.7 Most firms in the Eduventures’ survey allow employees to become eligible for the program after six months of service; rarely do eligibility requirements exceed one year. Twenty percent of firms in the survey impose service requirements after participation. Service requirements after participation are more common in plans that have unlimited tuition reimbursement. The survey also reports that over 90 percent of programs have a minimum grade standard for reimbursement, typically set at a “C” or better. Many companies tie grades directly to reimbursement percentages, making the cost of participation higher for workers who receive lower grades. The tuition reimbursement program offered at the case study insti5. Section 127 of the Internal Revenue Code outlines the tax-advantaged status of educational assistance plans provided by employers: “Gross income of an employee does not include amounts paid or expenses incurred by the employer for educational assistance to the employee if the assistance is furnished pursuant to a [educational assistance] program” (26 U.S.C.§ 127). 6. The survey was sponsored by Cenquest, a provider of managed education solutions, which helps companies create and manage tuition assistance programs (http://www .cenquest.com). Eduventures, who conducted the survey, is an independent research and advisory firm of corporate, postsecondary, and pre–K-12 learning markets (http://www .eduventures.com). 7. A common reaction to tuition reimbursement programs is to only attribute their provision by firms to their tax-advantaged status. However, this overlooks the trade-off between wage and nonwage compensation. If total compensation reflects the value of a worker’s marginal product, then benefits and wages are substitutes at the margin. Firms offer tuition reimbursement program instead of additional wages or other benefits if tuition benefits are more effective at attracting or retaining a certain type of worker. The tax-advantaged status of tuition reimbursement programs increases the value of these benefits to a worker facing a positive tax rate, but it cannot explain the effect of these programs on recruitment or turnover.

The Effect of Tuition Reimbursement on Turnover Table 6.1

201

Distribution of maximum reimbursements for tuition programs Amount